🤖 Robotics arXiv Digest

Monday, October 19, 2026

📄 30 papers 📂 6 research areas Generated by Claude

🔭 Research Landscape

The dominant theme today is what a VLA should predict before it acts, and the answers are converging on compact, structured intermediates rather than raw futures. VLDA (#13) argues that full future-state prediction wastes capacity on static appearance and instead learns a compact "Delta" of task-relevant change; TrajWAM (#19) makes a related bet for dynamic scenes by grounding imagination in forecast object-mask trajectories and distilling to one-step inference; PhysWord (#17) anchors latent dynamics to atomic action verbs; and LECDrive (#24) injects dense DINOv3/SAM3/DepthAnything3 knowledge through latent tokens. Ropedia W0 (#8) pushes the same idea to scale, with 10k hours of egocentric human data and 3k hours of robot data in a shared kinematic representation. The common move is to put a structured, low-bandwidth interface between perception and action.

A second thread is the efficiency and embodiment cost of deploying those models. SPAR (#28) shows that 96% of a pi0.5 language backbone can be pruned and recovered with RL, PaceVLA (#29) decides per step how coarsely to quantize actions in time, and HAF-VLA (#6) adapts a flow-matching VLA to humanoids by denoising locomotion, waist and arms in successive stages. Meanwhile the policy-learning papers attack data quality instead of model size: PGPG (#16) and ALIGN (#25) extract signal from imperfect rollouts without human labels or noisy thresholds, DISCO (#23) distills MCTS into a deployable policy, and Policy Search in Code Space (#5) lets coding agents write interpretable multi-agent pathfinding algorithms that beat competition entries by 35-59%. PhysRSI-eval (#27) is the measurement counterpart, asking how quickly and how far VLM agents improve from physical feedback.

Third, a cluster of papers is a corrective on evaluation and assumptions. ForagerSim (#2) finds learned policies lag humans once maps and topology get large, ReInsVLN (#18) shows that re-reading history recovers failures a frozen navigator already had the information to avoid, VeSTA (#20) shows that permissive simulator contact hides a physical executability gap in VLN, and HandsOn (#3) reports that the best trained multi-task dexterous policy reaches only 29.1% on Core-50 while GPT-6 prompting manages 5.2%. In estimation, Self-Supervised Inertial Odometry (#1) and the wheel-uncertainty LIO (#15) both replace hand-set noise or ground truth with learned structure, within 0.07 m/s of a supervised baseline in the first case. Taken together, the batch suggests the field is moving from "can the model act" toward "does the interface, the benchmark and the hardware assumption actually hold".

VLA & World-Action Models

How to make VLAs plan, predict and run faster: delta/trajectory-conditioned world-action models, structured latent reasoning, dense driving cognition, humanoid action flow, pruning and adaptive action chunking.

#6 HAF-VLA: Hierarchical Action Flow Framework for Huma...
#8 Ropedia W0: Scaling Loco-Manipulation World Models a...
#13 Vision-Language-Delta-Action: Learning Compact Delta...
#17 Physical Words: A Vocabulary for Structuring Latent...
#19 TrajWAM: An Efficient Trajectory-Guided World Action...
#24 Think Densely, Act Sparsely: Latent Expert Cognitive...
#28 SPAR: Structural Pruning and Action Recovery for Vis...
#29 PaceVLA: Adaptive Temporal Action Quantization in Vi...

Policy Learning, Search & Self-Improvement

Getting better policies from less: coding-agent policy search, sim-prior adaptation, preference and advantage conditioning from messy data, distilled tree search, and measuring physical self-improvement.

#5 Policy Search in Code Space for Multi-Agent Pathfind...
#10 Simulation Priors for Rapid Adaptation in Model-Base...
#14 Robot Learning with Vision-Language-Action Models fo...
#16 Progress-Guided Preference Generation for Aligning G...
#23 Long-Horizon Multi-Object Rearrangement via Distille...
#25 ALIGN: Aligning Condition Estimation and Expression...
#27 Where is Physical RSI capability from? An Observatio...

Embodied Navigation & Map Reasoning

Benchmarks and methods that expose what navigation agents actually lack: map-based reasoning, usable history, and physically executable trajectories.

#2 ForagerSim: Evaluating Embodied Map-Based Reasoning...
#18 ReInsVLN: Revisiting Historical Evidence for Trainin...
#20 Revealing and Bridging the Physical Executability Ga...

Humanoid, Legged & Aerial Platforms

Hardware-and-control work on new morphologies: supernumerary humanoid limbs, dual-humanoid interaction, thruster-assisted legged-aerial trotting, NAO imitation and multirotor wind sensing.

#4 Vehicle-State-Aware Confidence Estimation for Direct...
#9 Data-Efficient Learning from Demonstration for Human...
#21 Bio-inspired Thrust-Assisted Trotting on a legged-Ae...
#22 OmniInteract: Towards Unified Whole-Body Interaction...
#26 Humanoid SRL: Augmenting Humanoids with Supernumerar...

State Estimation, Mapping & Perception

Making estimators robust where classical assumptions fail: label-free inertial odometry, learned wheel uncertainty, dynamic-object SLAM and LiDAR-visual radiance voxels at campus scale.

#1 Self-Supervised Inertial Odometry
#11 CAOS-SLAM: Class-Agnostic Object-Aware Real-Time vSL...
#15 Learning Wheel Motion Uncertainty for Degeneracy-Awa...
#30 SpaRV: Sparse Radiance Voxels for Scalable Geometric...

Dexterous & Tactile Manipulation

Dexterous-hand infrastructure and sensing: a cross-embodiment benchmark where policies reach under 30%, object-aware retargeting, and a biomimetic skin to test receptor layouts.

#3 HandsOn: A Large-Scale Simulation Benchmark for Gene...
#7 DexOAK: Object-Aware Kinematic Retargeting
#12 Robotic Physical Comparison of Biologically Motivate...

VLA & World-Action Models

#6 h=29

HAF-VLA: Hierarchical Action Flow Framework for Humanoid Vision-Language-Action Control

2026-10-19 cs.RO 👤 Zhengping Che (h=29)

Langzhe Gu, Chengkai Hou, Meng Li, Xinhua Wang, Xinyuan Lv

Core Contributions

Addresses why generalist VLAs struggle on humanoids: base and waist errors trigger aggressive upper-body compensation in a high-dimensional whole-body action space.
Instead of one-shot chunk generation, a pretrained flow-matching VLA denoises a shared action space in stages: locomotion and head, then waist, then full-body manipulation.
Learnable stage embeddings plus cross-stage KV-cache conditioning let later stages refine what earlier stages committed to, keeping structural dependencies.
Reports significant gains over existing models on six real-world humanoid loco-manipulation tasks, a recipe for adapting mobile-manipulation VLAs to legged bodies.

Show abstract

Humanoid robots hold great promise as general-purpose agents in human-centered environments, yet adapting generalist vision-language-action (VLA) policies to real-world humanoid loco-manipulation remains challenging. Unlike fixed-base or wheeled manipulators, humanoids must coordinate locomotion, posture, gaze, and dual-arm manipulation over a high-dimensional whole-body action space, where inaccurate base or waist motions can easily induce unstable and aggressive upper-body compensation. We propose HAF-VLA, a hierarchical action-flow framework for humanoid VLA control. Built upon a pretrained flow-matching VLA, HAF-VLA inherits the vision-language prefix encoder and action expert weights, and adapts them to humanoid whole-body loco-manipulation with learnable stage embeddings and cross-stage KV-cache conditioning. Instead of generating the full action chunk in a single pass, HAF-VLA progressively denoises a shared whole-body action space: first locomotion and head commands, then waist refinement, and finally full-body manipulation. To preserve structural dependencies across stages, we introduce a cross-stage KV cache conditioning mechanism, enabling subsequent action generation stages to refine and augment their outputs based on the action representations from preceding stages. Experimental results across six real-world tasks demonstrate that our framework significantly enhances the performance of existing models in humanoid loco-manipulation tasks, demonstrating an effective pathway for adapting strong mobile-manipulation VLAs to humanoid whole-body loco-manipulation.

#8 h=24

Ropedia W0: Scaling Loco-Manipulation World Models across Embodiments with Human Experience

2026-10-19 cs.CVcs.RO 👤 Zhao-Xi Chen (h=24)

Tianqi Liu, Hanfei Zhao, Qifeng Liang, Hui En Pang, Jiarui Xu

Core Contributions

Builds a world foundation model that jointly covers locomotion and manipulation, across humans and robots, from egocentric views.
Backs it with Xp-Data (10k hours of human and 3k hours of robot data with unified annotations) and a kinematic representation encoding locomotion via camera pose and manipulation via projected skeletons.
Reports state of the art on five cross-embodiment benchmarks and held-out human video, with performance improving as human pre-training data scales for both human and robot evaluation.
After distillation it supports real-time, minute-long rollouts for live VR interaction, and is applied to policy evaluation, planning and world-action modeling.

Show abstract

World models enable embodied agents to anticipate action consequences, yet jointly modeling locomotion and manipulation across diverse embodiments remains challenging due to heterogeneous action spaces and limited multimodal supervision. We present Ropedia W0, the first egocentric loco-manipulation world foundation model spanning humans and robots with diverse embodiments. We curate Xp-Data, comprising 10k hours of egocentric human data and 3k hours of robot data with unified multimodal annotations and systematic quality control. A unified kinematic representation encodes locomotion through camera poses and manipulation through projected skeletons, with robot skeletons combining embodiment-specific forward kinematics with extensive manual keypoint annotation and correction. We also introduce Xp-Teleop, a long-horizon loco-manipulation benchmark based on high-quality dexterous humanoid teleoperation data. Evaluations demonstrate state-of-the-art performance across five cross-embodiment benchmarks (DreamDojo, $Ψ_0$, $ω$-HOME, Xp-Teleop, and AgiBot) and on held-out human egocentric video, from zero-shot inference to few-shot and full-data post-training. Scaling human pre-training data improves performance on both human and robot evaluations. After distillation, our model supports real-time, minute-long rollouts, enabling live VR interaction. Applications in policy evaluation, model-based planning, and world-action modeling further demonstrate the model's utility for embodied decision-making.

#13 h=18

Vision-Language-Delta-Action: Learning Compact Deltas as Actionable Transition Interfaces for Robotic Manipulation

2026-10-19 cs.ROcs.CV 👤 Rui Shao (h=18)

Lingsen Zhang, Wei Li, Jie He, Henghua Zhang, Kai Mo

Core Contributions

Argues that manipulation is realizing a state change, so full future prediction wastes capacity on static appearance; VLDA predicts a compact latent Delta as the interface between language-vision understanding and action.
Delta is trained with a Local Transition Filter (task-relevant changes, suppressing static reconstruction bias) and a Global Transition Anchor (coarse direction of change).
Delta-Driven Action Refinement then uses Delta as a structured seed for action denoising, grounding it in the action manifold.
Claims improved long-horizon consistency, robustness and generalization, implemented in the open-source LoongForge framework.

Show abstract

Robotic manipulation is not merely observation-to-action generation, but the realization of task-specified state changes in the physical world. Existing vision-language-action (VLA) methods directly learn observation-to-action mappings, imposing a substantial learning burden due to the gap between semantic understanding and action execution. Although world-action models (WAMs) improve control through future prediction, modeling complete future states wastes capacity on static appearance and action-irrelevant changes. We propose Vision-Language-Delta-Action (VLDA), a paradigm that reformulates full future prediction as compact actionable transition modeling. Its core is a compact Delta variable that captures latent world-state changes, serving as an interface between vision-language understanding and action generation. VLDA learns Delta through a unified objective composed of the Local Transition Filter (LT-Filter) and Global Transition Anchor (GT-Anchor). LT-Filter captures task-relevant manipulation-centric changes while suppressing static reconstruction biases. GT-Anchor provides coarse future-level regularization to preserve the overall direction of state changes. Delta-Driven Action Refinement (DDAR) further grounds Delta in the action manifold and injects it as a structured seed to guide action denoising. We implement VLDA within LoongForge, an open-source training framework for large-scale foundation and embodied models, using its modular and distributed infrastructure for unified two-stage optimization. Experiments show that VLDA improves long-horizon consistency, robustness, and generalization, validating actionable Delta as an alternative to direct action prediction and full future-state modeling.

#17 h=15

Physical Words: A Vocabulary for Structuring Latent Reasoning in Robot Manipulation

2026-10-19 cs.RO 👤 Soha Pouya (h=15)

Jiaming Liu, Xinghao Zhu, Yiming Zhang, Shanghang Zhang, Yinxi Wang

Core Contributions

Introduces Physical Words: atomic action verbs used as semantic anchors for continuous latent representations of physical dynamics, between language reasoning (too coarse) and unstructured latent prediction.
PhysWord predicts upcoming action words and aligns their hidden states with latent targets from future observations, building a temporally structured space with visual and optional tactile variants.
Reports 90.3% on visuotactile UniVTAC and 99.6% on LIBERO-Long, and wins on real bimanual-gripper and dexterous-hand tasks.
Gives reasoning-before-acting models reusable primitives for long-horizon manipulation.

Show abstract

Multimodal reasoning provides robotic foundation models with predictive context needed for reliable action generation. However, language-based reasoning captures task semantics but struggles to express fine-grained physical dynamics, while latent reasoning models future state evolution but offers limited structure for organizing these dynamics into reusable reasoning primitives. To bridge this gap, we introduce PhysWord, a reasoning-before-acting foundation model built on Physical Words, which use atomic action verbs as shared semantic anchors for continuous representations of physical dynamics. Specifically, PhysWord predicts upcoming atomic action words and directly aligns their contextual hidden states with latent targets derived from the corresponding future observations. Building on these primitives, we construct a temporally structured PhysWord space with visual and optional tactile variants of each action word. Autoregressively predicting Physical Words across future timesteps captures transitions between execution stages, providing structured multimodal context for action generation. PhysWord achieves state-of-the-art success rates of 90.3\% on visuotactile UniVTAC and 99.6\% on vision-based LIBERO-Long, and outperforms competing methods on real-world tasks using bimanual grippers and dexterous hands. Further evaluations demonstrate that our structured latent reasoning framework generalizes to unseen scenarios and enables long-horizon manipulation.

#19 h=14

TrajWAM: An Efficient Trajectory-Guided World Action Model for Dynamic Robotic Manipulation

2026-10-19 cs.RO 👤 Jian-Min Ji (h=14)

Haohan Geng, Sha Zhang, Weishuo Sun, Guanting Ye, Wenhao Yu

Core Contributions

Targets dynamic object manipulation where scene-level video prediction underrepresents the target and iterative denoising is too slow.
A history-conditioned predictor forecasts future object masks; a lightweight Trajectory ControlNet steers a pretrained video model to trajectory-consistent futures that condition the action expert.
Decoupled distillation (distribution matching for video, ControlNet realignment, action rollout distillation) enables one-step video-action inference while keeping the trajectory pathway.
Claims state-of-the-art dynamic manipulation in simulation and real world with far lower latency than iterative world-action models.

Show abstract

Dynamic object manipulation requires robots to anticipate target motion while replanning fast enough to avoid executing outdated actions. Existing approaches either rely on fast reactive control without an explicit future target state or encode anticipated motion only implicitly. World Action Models (WAMs) offer a promising alternative by coupling future visual prediction with action generation, but scene-level video objectives may underrepresent target dynamics that are critical for manipulation, while iterative video-action denoising introduces substantial latency. We present TrajWAM, a trajectory-guided world action model for efficient dynamic robotic manipulation. Our key insight is to explicitly ground future visual imagination in predicted target-object motion and use the resulting future representation to condition action generation, while preserving this anticipation pathway after acceleration. Given a short history of object masks, a history-conditioned motion predictor forecasts future masks encoding the target object's trajectory. A lightweight Trajectory ControlNet then steers a pretrained video model toward trajectory-consistent futures. The resulting future representation conditions an action expert to predict executable action chunks. To enable one-step inference without breaking this interface, we introduce a decoupled distillation scheme that combines distribution matching for the video backbone, Trajectory ControlNet realignment, and action rollout distillation to support accurate one-step action prediction conditioned on the compressed future representation. Comprehensive evaluations in simulation and the real world show that TrajWAM achieves state-of-the-art performance on dynamic object manipulation tasks while enabling efficient one-step video-action inference. Project page: https://kevingeng3.github.io/TrajWAM/.

#24 h=9

Think Densely, Act Sparsely: Latent Expert Cognitive Chains for Vision-Language-Action Autonomous Driving

2026-10-19 cs.CVcs.RO 👤 Hang-Jun Ye (h=9)

Jie Wang, Guang Li, Zhijian Huang, Jinlong Li, Chenxu Dang

Core Contributions

Argues driving VLAs fail not for lack of capacity but because sparse supervision (language, trajectories) never forces a dense world representation; LECDrive adds task-specific latent tokens that absorb expert knowledge.
Tokens internalize DINOv3 visual features, SAM3 semantics and DepthAnything3 geometry through a curriculum from representation to semantics to geometry, keeping inference efficient.
Consistently improves base VLMs across driving benchmarks for perception, state prediction and trajectory planning.
Distills heavyweight vision foundation models into the VLA without running them at test time.

Show abstract

Autonomous driving fundamentally demands dense understanding of visual, semantic, and geometric information, yet existing Vision-Language-Action (VLA) models are typically trained with sparse supervision such as language instructions or trajectory signals. This direct mapping from dense visual observations to sparse representations forces the model to compress or even discard critical scene information essential for safe driving. We argue that this limitation does not stem from insufficient model capacity, but from the lack of explicit mechanisms that encourage the formation of dense world representations in the latent space. In this spirit, we propose a novel paradigm for driving intelligence, Think Densely, Act Sparsely, and instantiate it with an efficient and interpretable framework, LECDrive (Latent Expert Cognitive Chains for Vision-Language-Action autonomous Driving). The core idea is to mimic the hierarchical perception process of humans by embedding a progressive chain of latent expert cognition within the VLA model. Concretely, we inject a compact set of task-specific latent tokens as carriers, enabling the model to internalize dense knowledge from foundational vision models, including visual representations (DINOv3), semantic structures (SAM3), and spatial geometry (DepthAnything3), while maintaining efficient inference. This interdependent cognitive chain follows a curriculum learning principle, where supervision is progressively structured from representation to semantics and geometry. Extensive experiments demonstrate that LECDrive consistently improves the performance of base VLMs across multiple autonomous driving benchmarks, covering key tasks such as object perception, state prediction, and trajectory planning. These findings suggest that constructing dense world models through structured latent expert chains is a promising direction for VLA-based autonomous driving.

#28 h=7

SPAR: Structural Pruning and Action Recovery for Vision-Language-Action Models

2026-10-19 cs.ROcs.AI 👤 Yuan Du (h=7)

Shenli Zheng, Zhou Fang, Yuan Du, Yijiang Liu, Li Du

Core Contributions

Prunes the hidden and feed-forward dimensions of the LLM backbone with pruning-aware tuning, then uses RL to recover manipulation performance, which supervised fine-tuning cannot fully do.
On pi0.5, the language backbone shrinks from 2.34B to 0.09B (about 96%) and the full model from 3.37B to 1.12B (about 67%).
RL raises average LIBERO success from 84.55% to 93.45%, an 8.90-point gain, with 1.98x faster full-policy inference.
Indicates large redundancy in VLA intermediate dimensions and a route to onboard deployment.

Show abstract

The limited onboard computing resources of robots make the inference demands of vision-language-action (VLA) models costly to accommodate. Reducing these demands through aggressive compression, however, can degrade manipulation performance beyond what supervised fine-tuning alone recovers. We present SPAR (Structural Pruning and Action Recovery), which simultaneously prunes the hidden and feed-forward intermediate dimensions of the large language model (LLM) backbone through pruning-aware tuning, then applies reinforcement learning (RL) to recover policy performance. Applied to $π_{0.5}$, SPAR reduces the language backbone from 2.34B to 0.09B parameters and the full model from 3.37B to 1.12B, removing approximately 96% and 67% of their parameters, respectively. At this compression level, RL raises average success across the four LIBERO suites from 84.55% to 93.45%, an 8.90-percentage-point gain. The compressed policy achieves a $1.98\times$ full-policy inference speedup over the unpruned model. These results identify substantial intermediate-dimension redundancy and show that post-pruning RL can recover manipulation performance in compact policies.

#29 h=7

PaceVLA: Adaptive Temporal Action Quantization in Vision-Language-Action Models

2026-10-19 cs.ROcs.AI 👤 Joonwoo Ahn (h=7)

Hojin Jeon, Jimin Lee, Taekwan Kim, Nayoung Oh, Taeyoung Kim

Core Contributions

Targets slow execution from fine-grained teleoperation actions; naive temporal action quantization trades speed for accuracy in precision-critical phases.
PaceVLA uses mode-specific action decoders on a shared action expert, supporting several temporal granularities and prediction horizons.
A quantizability-guided scheme and a context-conditioned router choose coarse or fine execution at each inference step, speeding up easy phases and slowing for delicate ones.
Improves efficiency while maintaining success across three simulation benchmarks and six real tasks, including long-horizon and dexterous ones.

Show abstract

Fast and reliable manipulation is crucial for practical robot deployment. Although Vision-Language-Action (VLA) models have recently shown remarkable progress as generalist robot policies, their execution often remains temporally inefficient. A key source of this inefficiency is that policies learn fine-grained temporal patterns from teleoperated demonstrations, requiring many low-level action steps. Temporal action quantization offers a simple way to accelerate execution by aggregating consecutive low-level actions into temporally coarser actions. However, applying action quantization without considering when fine-grained control is required introduces a speed-accuracy trade-off: accelerated actions may efficiently reduce redundant steps in some phases, while discarding critical motion details in precision-sensitive interactions. To address this, we propose PaceVLA, a unified framework for adaptive temporal execution in VLAs that integrates two coupled components: (1) mode-specific action decoders built on a shared action expert, supporting multiple temporal granularities and prediction horizons, and (2) a quantizability-guided mode selection scheme which decides whether coarse execution is appropriate, and a context-conditioned router that selects an execution mode at each inference step. Experiments across three multi-task simulation benchmarks and six real-world tasks show that PaceVLA improves execution efficiency while maintaining task success rate across multiple embodiments, including long-horizon and dexterous manipulation.

Policy Learning, Search & Self-Improvement

#5 h=32

Policy Search in Code Space for Multi-Agent Pathfinding

2026-10-19 cs.ROcs.AI 👤 V. Farias (h=32)

Matthias Eder, Pantelis Emmanouil, Vivek Farias, Fabian Fischer, Adam Jozefiak

Core Contributions

Proposes searching directly in code space: coding agents propose, implement, evaluate, diagnose and refine multi-agent coordination algorithms under human direction, with explicit correctness verification.
On a hard 2024 League of Robot Runners map, a two-day search yielded a policy beating the stronger locally reproduced entry by 35.0% on the densest instance and generalizing to unseen maps.
In simulated warehouses with irregular lanes and asynchronous motion, a week-long search found a new class of MAPF algorithm with 58.9% higher throughput than a calibrated proprietary solution.
Unlike RL, the output is readable code, and the authors prove under stated assumptions that it is collision-free and completes every accepted goal in finite time.

Show abstract

Developing coordination policies for complex multi-agent systems relies on either labor-intensive algorithm design or opaque task-specific reinforcement learning. With recent advances in autonomous coding agents, we show that policy search directly in code space offers a compelling third path that combines automated optimization with explicit interpretability. We investigate a human-directed workflow in which coding agents iteratively propose, implement, evaluate, diagnose, and eventually refine algorithms guided by simulation feedback and explicit correctness verification. We conducted two case studies and found that this workflow robustly outperformed any manually engineered solutions. On a challenging map in the 2024 League of Robot Runners competition, a two-day search discovered a policy that exceeds the stronger locally reproduced competition entry by 35.0% on the densest instance and generalizes to unseen map instances. In simulated real-world warehouses with irregular lanes, continuous robot geometry, and asynchronous motion, a week-long search discovered an entirely new class of multi-agent pathfinding algorithm that outperforms a calibrated proprietary solution in throughput by 58.9%. Under stated assumptions, we further prove that the new algorithm is collision-free and completes, in finite time, every accepted goal that is not withdrawn or replaced. Together, the studies show that policy search in code space powered by state-of-the-art coding agents offers a highly competitive path to building real engineering solutions that are both effective and analytically tractable.

#10 h=21

Simulation Priors for Rapid Adaptation in Model-Based RL for Autonomous Racing

2026-10-19 cs.LGcs.ROeess.SY 👤 Nicklas Hansen (h=21)

Adrian Remonda, Jiajun Xi, Nicklas Hansen, Marcus Greiff, Thomas Lew

Core Contributions

Treats simulators as biased priors rather than transfer targets: PIA pretrains a world model on mixed low- and high-fidelity simulated trajectories, then corrects the mismatch online with uncertainty-regularized planning.
In simulation it adapts faster than strong baselines across track and surface shifts, with lower lap times and fewer failures.
On a real full-size autonomous race car, it adapts few-shot with under one hour of on-vehicle interaction, which matters where data is costly and failures are unsafe.
Shows cheap, imperfect simulation can be useful without system identification of an accurate simulator.

Show abstract

Reinforcement learning enables learning high-performance control policies, but its reliance on extensive interaction limits deployment on physical systems where data collection is costly, and failures are unsafe. We introduce Prior-Informed Adaptation (PIA), an offline-to-online model-based RL framework that uses structured simulation priors for rapid adaptation. PIA pretrains a world model on heterogeneous low- and high-fidelity simulated trajectories, then adapts it with limited target-domain interaction using uncertainty-regularized planning. Rather than requiring an accurate simulator for policy transfer, PIA treats simulated trajectories as biased priors whose residual mismatch is corrected during online adaptation. We evaluate PIA in autonomous racing using high-fidelity simulation and on a real full-size autonomous race car. In simulation, PIA adapts faster across track and surface shifts than strong baselines, achieving lower lap times and fewer failures. On hardware, PIA achieves reliable few-shot adaptation with less than one hour of on-vehicle interaction.

#14 h=17

Robot Learning with Vision-Language-Action Models for Agricultural Manipulation

2026-10-19 cs.RO 👤 J. M. Velni (h=17)

Hamid Hafezi, Javad Mohammadpour Velni, Mohammadreza Davoodi

Core Contributions

Builds a sim-to-real pipeline for tomato-plant reaching and grasping: a MuJoCo environment with varied plant structures, joystick-teleoperated demonstrations, then real Franka fine-tuning.
Compares two pretrained generalists, Octo and pi0.5, under in-distribution, viewpoint/spatial-shift and out-of-distribution conditions.
pi0.5 scores higher and degrades less under distribution shift; the paper also studies how much proprioception contributes.
Practical evidence that simulated demonstrations can cheaply seed generalist policies for agriculture, where real data is expensive.

Show abstract

Robotic manipulation in agricultural environments remains challenging due to variations in plant geometry, occlusion, and camera viewpoints. Pretrained vision-language-action (VLA) and generalist robot policies offer a promising approach to addressing these challenges by learning manipulation behaviors from multimodal observations and language instructions. However, adapting these policies to agricultural tasks requires domain-specific demonstrations, which are costly and time-consuming to collect on physical robots. In this work, we develop a simulation-to-real robot-learning framework for agricultural reaching and grasping that integrates task-specific demonstration collection and pretrained policy adaptation. A MuJoCo-based agricultural environment is developed with diverse tomato plant structures to support demonstration collection and controlled evaluation. Expert demonstrations are collected through joystick teleoperation, with a low-level controller used to execute robot actions. Two pretrained generalist robot policies, Octo and $π_{0.5}$, are adapted using simulated demonstrations, followed by further fine-tuning on a physical Franka Research 3 manipulator using real-robot demonstrations. The adapted policies are evaluated under in-distribution, viewpoint/spatial-shift, and out-of-distribution conditions in simulation, followed by physical-robot experiments. The influence of proprioceptive information on policy performance is also investigated. Results show that $π_{0.5}$ achieves higher task-performance scores than Octo and exhibits less degradation under distribution shifts. These findings demonstrate the potential of integrating pretrained generalist policies with task-specific demonstration learning for agricultural manipulation.

#16 h=17

Progress-Guided Preference Generation for Aligning Generative Control Policies

2026-10-19 cs.LGcs.AIcs.RO 👤 V. Ngo (h=17)

Tung M. Luu, Tri Ton, Duc M. Nguyen, Xuan Thanh Nguyen, Nguyen Gia Binh

Core Contributions

Replaces human preference labels with automatic ones: ordered coverage alignment measures how much of the expert demonstrations' sequential progress each rollout achieves.
Ranks autonomously collected rollouts by progress to produce pseudo-preferences, with enhancements for label reliability, then fine-tunes pretrained generative policies.
Outperforms supervised fine-tuning on long-horizon manipulation and works with both diffusion and flow-matching policies.
Removes both the human annotation and the hand-designed reward typically needed for preference alignment.

Show abstract

Generative policies have achieved strong performance in imitation learning tasks, but their success is often constrained by the limited availability of expert demonstrations. Although preference-based fine-tuning can leverage suboptimal rollouts to improve performance, existing approaches heavily rely on expensive human annotations. In this work, we introduce Progress-Guided Preference Generation (PGPG), an unsupervised preference alignment framework that eliminates the need for human-in-the-loop supervision. Our method leverages ordered coverage alignment to measure sequential progress relative to a small set of expert demonstrations, enabling the automatic extraction of pseudo-preference labels from autonomously collected rollouts. Furthermore, we introduce key enhancements to improve the reliability of these progress-guided preferences, enabling effective fine-tuning of pretrained generative policies without requiring manually designed reward functions. Extensive experiments across a suite of challenging long-horizon manipulation tasks demonstrate that PGPG outperforms standard supervised fine-tuning baselines. Moreover, our conceptually simple framework is broadly applicable across multiple classes of generative policies, demonstrating strong versatility when integrated with both diffusion and flow-matching policies.

#23 h=9

Long-Horizon Multi-Object Rearrangement via Distilled Tree Search over Continuous Actions

2026-10-19 cs.RO 👤 Maxim Likhachev (h=9)

Kallol Saha, Yining Hong, Amber Li, Yejin Choi, Maxim Likhachev

Core Contributions

Asks how to transfer long-horizon search into policies in continuous action spaces: progressive-widening MCTS in GPU-parallel physics simulation, with a learned policy proposing actions and a learned value function scoring progress.
Periodically distills the MCTS plans back into the policy-value model, so later searches start from better proposals and values (an AlphaZero-style loop for continuous manipulation).
At deployment the policy-value model runs without simulator search.
Outperforms behavior cloning, RL and search baselines on multi-object rearrangement in simulation and the real world.

Show abstract

Despite recent advances in policy learning, long-horizon manipulation remains difficult because learned policies must avoid compounding errors while preserving future feasibility. While search-based planning can explicitly reason over such future consequences, it quickly becomes expensive in high-dimensional continuous action spaces. This raises a central question: how can long-horizon reasoning from search be transferred into policies in continuous action spaces? To this end, we propose DISCO: Distilled Tree Search over Continuous Actions. DISCO uses progressive-widening MCTS in physics simulation to generate long-horizon plans. During search, a learned policy proposes continuous actions, while a learned value function scores actions by their expected progress toward task completion. Using GPU-parallelized simulation, DISCO periodically distills MCTS-generated plans into the policy-value model, improving the action proposals and value estimates used in subsequent searches. Finally, the learned policy-value model can be directly deployed without the need for simulator-based search at test time. Across simulated and real-world rearrangement tasks, DISCO outperforms behavior cloning, reinforcement learning, and search-based baselines.

#25 h=8

ALIGN: Aligning Condition Estimation and Expression in Advantage-Conditioned Behavior Cloning

2026-10-19 cs.ROcs.AI 👤 Zechen Wang (h=8)

Zechen Wang, Zhenqi Qiu, Houlong Xiong, Suohang Zhang, Yiyu Ren

Core Contributions

Diagnoses two weaknesses in advantage-conditioned behavior cloning on mixed-quality data: thresholding advantage scores gives noisy conditions, and prompt-based conditioning only weakly affects the action vector field.
ALIGN predicts categorical pairwise conditions directly in the policy-facing label space and injects them into a DiT action expert via AdaRMS modulation fused with flow time.
On retry-heavy real datasets, closed-loop success is 90.0% (ARX X5 bowl stacking) and 56.7% (Franka key hanging) versus 56.7% and 33.3% for the best AC-BC baseline.
Lets teams keep recovery and failure data, which usually gets filtered out.

Show abstract

Human-collected robot manipulation data often mix successful behavior with retries, pauses, recoveries, and failures. Advantage-Conditioned Behavior Cloning (AC-BC) can exploit such mixed-quality data, but only if the conditioning signal is both accurately constructed from offline trajectories and effectively incorporated into the policy's action-vector-field prediction. Existing pipelines can weaken this signal by thresholding estimator-derived advantage scores into noisy conditions, and by routing conditions through prompts that weakly affect the action vector field. We introduce ALIGN, a condition-aligned AC-BC framework for vision-language-action policies. ALIGN directly predicts categorical pairwise conditions in the policy-facing label space and injects them into a DiT action expert through AdaRMS modulation fused with flow time. We evaluate ALIGN on deliberately mixed real-robot datasets for ARX X5 bowl stacking and Franka key hanging, each combining standard demonstrations with retry-heavy trajectories. ALIGN improves condition fidelity, strengthens conditional vector-field separability, and achieves 90.0% and 56.7% closed-loop success, compared with 56.7% and 33.3% for the strongest AC-BC baseline. These results indicate that aligning conditions from estimation to action prediction is an effective way to extract robust manipulation policies from data in which success, recovery and error coexist.

#27 h=7

Where is Physical RSI capability from? An Observational Scaling Law of LLM-as-policy

2026-10-19 cs.ROcs.AIcs.CV 👤 Ze-Kun Qi (h=7)

Zekai Li, Jiayi Yu, Jinlu Zhang, Luyang Niu, Zekun Qi

Core Contributions

Defines PhysRSI-eval, a five-axis metric (first attempt, efficiency, success, budget, improvement from failure) for physical recursive self-improvement, run on 18 manipulation tasks with a fixed interaction budget.
Proposes a Physical Barrel Hypothesis: different capabilities limit different parts of improvement.
The metrics separate systems with similar public benchmark averages, and high scores need not come with early or efficient gains.
Observational scaling predicts VLM-agent performance from public capabilities and forecasts later progress from early feedback; workspace interventions show scene-specific adaptation and reusable experience can coexist.

Show abstract

In this paper, we systematically investigate the mechanisms, scaling behavior, and generalization of physical Recursive self-improvement. We introduce PhysRSI-eval to quantify physical RSI through a five-axis metric: First-Attempt Performance, Efficiency, Success, Budget and Improvement from failure. Different VLMs are evaluated on 18 manipulation tasks under a fixed interaction budget. The analysis of experimental results motivates the Physical Barrel Hypothesis: different capabilities may constrain different parts of improvement. The five-axis metrics also distinguish systems with close public benchmark averages and show that high scores need not come with early or efficient gains. We study observational scaling from two perspectives: predicting the relative performance of VLM agents from their public capabilities, and forecasting later progress from early physical feedback. Together, these analyses ask how quickly and how far an agent can improve. Moreover, we conduct workspace interventions and then analyze what improvement leaves behind. Removing history reduces late-round success, frozen workspaces remain useful on unseen layouts, and three cross-task cases achieve success earlier. The results suggest that scene-specific adaptation and reusable experience can coexist.

Embodied Navigation & Map Reasoning

#2 h=45

ForagerSim: Evaluating Embodied Map-Based Reasoning in Large-Scale Dynamic Worlds

2026-10-19 cs.ROcs.LG 👤 M. Saska (h=45)

Tomáš Musil, Martin Kotík, Michal Vavrečka, Karla Štěpánová, Sebastian Scherer

Core Contributions

Argues that existing embodied benchmarks rarely force map-based reasoning, so ForagerSim builds scenarios from room-scale mazes to multi-km open worlds with independent difficulty knobs for perception, scale, topology, dynamics (smoke, debris, sensor failures) and task.
Draws on animal foraging (e.g. well-alternation) so the same environment can test cognitive-map hypotheses and robotics search-and-rescue.
Result across 8 worlds: with simplified 2D sensing, RL does well on multi-object search in simple topologies, but falls short of a human player in large or complex ones where an engineered map-based stack still works, pointing at map-building as the bottleneck.
Gives researchers a controlled way to isolate which axis breaks learned agents instead of reporting one aggregate success rate.

Show abstract

Animals build cognitive maps of vast environments, use them to perform map-based navigation, and might even be reusing the same circuitry used in spatial reasoning for creating and using maps of arbitrary task-spaces. In contrast, current embodied AI systems either rely on hand-engineered mapping pipelines, which break in the presence of unmodeled dynamics and are restricted to metric space, or on learned policies that rarely form map-like representations at all. Developing architectures that could learn to build and use map-like representations as flexibly as animals is, however, difficult because few embodied AI benchmarks actually require map-based reasoning. To address this, we present ForagerSim --- an interdisciplinary research platform and simulation environment focused on embodied tasks that require map-based reasoning, with scenarios ranging from mobile-robot search and rescue to animal foraging in both complex mazes and vast, potentially infinite open areas. ForagerSim is highly modular and allows modifying independent axes of difficulty related to perception (abstract 2D sensing to full RGB or LiDAR), scale (room-scale to multi-km scale), topology (open-fields to complex mazes), dynamics (smoke, debris, sensorimotor failures and others) and tasks (from simple 2D coverage to foraging-inspired well-alternation tasks). We compare reinforcement learning baselines with a simple engineered robotics stack and a human player on 8 example worlds which show how different difficulty axes affect existing methods. Our results show that, when sensing is simplified to 2D, learned policies can perform multi-object search and well-alternation tasks surprisingly well in simple topologies, but they fall short of human performance in larger or more complex-topology environments where map-based baselines perform well, suggesting that map-based reasoning is the bottleneck in these scenarios.

#18 h=15

ReInsVLN: Revisiting Historical Evidence for Training-Free Vision-and-Language Navigation

2026-10-19 cs.RO 👤 Si-Wei Cheng (h=15)

Siwei Cheng, Hao Wu, Zitong Qi, Xinming Zhang, Haoyu Xi

Core Contributions

Introduces ReInsBench to test history use through controlled continuation trials, finding that prompting a navigator to review its own history recovers failures using existing context.
ReInsVLN, training-free, compresses history by removing temporal redundancy, protecting coverage and selecting tokens by instruction, then schedules forgetting-inspired reminders with no extra model calls.
On R2R Val-Unseen it lifts StreamVLN and InternVLA-N1 by 8.48 and 3.37 points to 64.27% and 64.44% success.
Cuts prefill latency by up to 21.77% and KV-cache memory by up to 32.25%, and works on real humanoid and quadruped robots.

Show abstract

Vision-and-language navigation (VLN) relies on retaining and using relevant observations. Yet growing histories introduce redundancy, and retained information may remain underused. We introduce ReInsBench, a diagnostic benchmark that examines history utilization through controlled continuation trials. Its results show that history-review reinstruction can recover navigation failures using the existing task context. Motivated by this finding, we propose ReInsVLN, a training-free framework supporting memory construction and utilization in frozen navigators. Coverage- and Instruction-Aware Compression constructs compact visual context through temporal redundancy reduction, coverage-aware protection, and instruction-guided token selection. Forgetting-Inspired History Reminders guide navigators to revisit retained observations according to a predefined schedule, without additional model calls. On R2R Val-Unseen, ReInsVLN improves success rates by 8.48 and 3.37 percentage points over the native StreamVLN and InternVLA-N1, reaching 64.27% and 64.44%, respectively. Visual compression largely preserves navigation performance while reducing prefill latency and KV cache memory by up to 21.77% and 32.25%, respectively, on matched observation workloads. Real-world trials on humanoid and quadruped robots further support the applicability of ReInsVLN to sequential instruction following and ordinal navigation.

#20 h=14

Revealing and Bridging the Physical Executability Gap in Vision-Language Navigation

2026-10-19 cs.RO 👤 Yu-Wing Tai (h=14)

Gang He, Zhile Cheng, Kepeng Xu, Zhenyang Liu, Yu-Wing Tai

Core Contributions

Shows that continuous VLN success can be inflated by permissive simulator contact that hides physically infeasible actions, a physical executability gap.
VeSTA aligns policies with executable futures: trajectory-relative group policy optimization uses strict-contact feedback, and a Traversability Risk Critic picks among multimodal candidate futures.
On R2R and RxR, existing agents depend heavily on simulator contact semantics; under strict contact VeSTA reduces collision-induced failures while preserving or improving goal-reaching.
Real-robot experiments suggest the learned traversability preference transfers, and the paper argues for stricter VLN evaluation.

Show abstract

Continuous vision-and-language navigation (VLN) is typically evaluated by goal-reaching success, assuming it reflects an agent's ability to execute language-grounded behavior. We show that this assumption can fail: permissive contact models in current simulators may mask physically infeasible actions, creating a physical executability gap between semantic competence and embodied realization. To bridge this gap, we introduce Visual-Semantic Traversability Alignment (VeSTA), a framework that aligns VLN policies with physically executable futures. We define visual-semantic traversability as the ability to predict instruction-conditioned trajectories that are semantically valid, geometrically feasible, and physically robust. VeSTA addresses this misalignment through two complementary mechanisms: improving executable support in trajectory generation and risk-aware selection among multimodal futures. Specifically, Trajectory Relative Group Policy Optimization uses strict-contact feedback to reshape the candidate distribution toward executable regions, while the Traversability Risk Critic learns a risk-adjusted scalar utility for candidate-wise future selection. Experiments on R2R and RxR reveal substantial dependence of existing VLN agents on simulator contact semantics. Under strict-contact evaluation, VeSTA substantially reduces collision-induced failures while preserving or improving goal-reaching performance, and controlled real-world robot experiments provide evidence that the learned preference transfers beyond simulation.

Humanoid, Legged & Aerial Platforms

#4 h=35

Vehicle-State-Aware Confidence Estimation for Direct Wind Sensing on a Multirotor

2026-10-19 cs.RO 👤 C. Yeow (h=35)

Sri Arjun Senthil Kumar, April Zi Qian Chan, Chen-Hua Yeow

Core Contributions

Treats the anemometer on a multirotor as a sensor whose trust depends on the vehicle's own motion, using horizontal and vertical speed to score each wind observation.
A two-feature confidence model reaches ROC-AUC 0.995 and flags 98.84% of translation observations as unreliable, validated on held-out autonomous waypoint flight.
Plugged into a 2D wind estimator, it cuts median gain during translation from 0.207 to 0.064 while keeping 0.151 in hover, so the estimator trusts data only when the airflow is clean.
Shows proprioception alone can gate atmospheric measurements, a cheap fix with no extra sensors.

Show abstract

Direct wind sensing from a multi-rotor UAV enables localized atmospheric measurements, but the reliability of the anemometer depends on vehicle motion. This work presents a vehicle-state-aware confidence method that uses horizontal and vertical vehicle speed to quantify the acquisition of each wind observation. The resulting confidence score is used to adapt measurement influence to a two-dimensional vector estimator. Experiments were conducted during horizontal translation, vertical motion, and held-out autonomous waypoint flight. The proposed two-feature formulation achieved an ROC-AUC of 0.995, with 98.84% of translation observations identified as unreliable. The confidence-aware adaptation reduced median estimator gain during translation from 0.207 to 0.064 while retaining median gain of 0.151 during hover. The results show that vehicle state provides useful proprioceptive context for regulating the influence of direct airborne wind observations.

#9 h=23

Data-Efficient Learning from Demonstration for Humanoid Robots via Feature-Level Perception and Prompt-Based Object Detection

2026-10-19 cs.RO 👤 M. T. Masouleh (h=23)

Keyhan Rayati, AliReza Beigy, Anoosheh Saadati, Pourya Shahverdi, Mehdi Tale Masouleh

Core Contributions

Avoids end-to-end image training by learning from compact features: Kinect joint estimates, prompt-based object detection and depth give human motion and 3D object locations.
A temporal network maps those features to NAO joint-angle trajectories, keeping data needs small for a low-cost humanoid.
78% pick-and-place success with only 16 demonstrations (86% with 100) and 71% on pushing with 8 demonstrations.
Useful as a data-efficiency reference point for imitation on constrained hardware, though limited to a simple platform and tasks.

Show abstract

Learning from Demonstration (LfD) allows robots to learn new skills from human examples; however, many vision-based methods require large datasets and resource-intensive end-to-end training. This paper introduces a data-efficient LfD framework for humanoid robots that learns manipulation skills from compact, task-specific features instead of raw images. The proposed approach integrates layered visual preprocessing, Kinect-based human joint estimation, prompt-based object detection, and depth information to extract human motion features and three-dimensional object locations from demonstration videos. These features are used to train a temporal deep learning model that predicts robot joint-angle trajectories for execution. The framework is implemented on a NAO humanoid robot, evaluated quantitatively in simulation, and demonstrated on physical hardware. In simulation, the robot achieves a 78% success rate on a pick-and-place task using only 16 demonstrations, which increases to 86% when trained with 100 demonstrations. The framework also achieves a 71% success rate on a pushing task using eight demonstrations, highlighting its effectiveness for data-efficient humanoid skill acquisition.

#21 h=13

Bio-inspired Thrust-Assisted Trotting on a legged-Aerial Robot via Real-Time QP-MPC

2026-10-19 cs.RO 👤 Eric Sihite (h=13)

Shreyansh Pitroda, Eric Sihite, Alireza Ramezani, Morteza Gharib

Core Contributions

Implements, on the Harpy thruster-assisted biped, a condensed QP-MPC that jointly regulates ground reaction force and thrust vectoring at 500 Hz, inspired by wing-assisted running in chukar partridges.
A decoupled design controls sagittal dynamics with MPC and frontal dynamics with cascade orientation stabilization.
The QP solves in 0.18 ms against a 2 ms budget while respecting the friction cone, with joint tracking within 1 degree.
Hardware trials at 0 and 0.1 m/s show mean velocity tracks reference despite oscillating instantaneous velocity, a first step for legged-aerial hybrids.

Show abstract

This paper presents the hardware implementation of a quadratic program-based controller on Harpy: thruster-assisted bipedal robot to perform thruster-assisted trotting. This is inspired by wing-assisted locomotion observed in birds such as the chukar partridge, which employ thrust vectoring to manipulate the ground reaction force. The proposed controller formulates the thruster-assisted spring-loaded inverted pendulum dynamics as a condensed QP-based model predictive control problem, enabling simultaneous regulation of ground contact force and active thrust vectoring at 500 Hz. A decoupled control architecture is used, where the robot's sagittal dynamics are controlled using QP-MPC while the frontal dynamics is regulated using a cascade orientation stabilization controller. Two experiments were conducted with different forward reference velocities of 0 m/s and 0.1 m/s. While the instantaneous velocity oscillates, the mean velocities can track their respective reference. Additionally, joint tracking error remains within 1 deg while tracking the desired leg length given by the leg trajectory controller and desired swing trajectory given by Raibert's heuristic. The QP-MPC problem is consistently solved with 0.18ms, which is well within the 2ms computation budget while satisfying the GRF friction cone constraint.

#22 h=10

OmniInteract: Towards Unified Whole-Body Interaction Control for Dual Humanoids

2026-10-19 cs.RO 👤 Yi-Lin Wei (h=10)

Yinghao Li, Tianyu Zhang, Zhuohao Li, Yi-Lin Wei, Fa-Ting Hong

Core Contributions

Tackles dual-humanoid physical interaction, where human-to-robot retargeting breaks contact and where learning skills and partner coordination jointly loses fidelity as diversity grows.
Surface Contact-Aware Retargeting uses dense human-robot surface correspondences to retarget both agents together, preserving contact interfaces and near-field geometry.
A Scalable Flow-based Interaction Policy keeps a frozen whole-body motion prior, learns partner-aware residual experts, and unifies them with conditional flow matching.
Improves retargeting fidelity and coordination as skills grow, with real-world contact establishment and perturbation recovery.

Show abstract

Physical interaction is a key capability for humanoid collaboration, requiring coordinated whole-body motion and precise contact. However, learning diverse dual-humanoid interactions faces two major challenges: (1) human-robot morphological differences disrupt contact interfaces and spatial relationships during retargeting, and (2) jointly learning motor skills and partner coordination makes it difficult to retain control fidelity as interaction diversity grows. In this paper, we propose OmniInteract to address these challenges through surface-aware modeling and scalable residual policy learning. First, we introduce Surface Contact-Aware Retargeting, which leverages dense human-robot surface correspondences to jointly retarget both humanoids, preserving explicit interaction contact interfaces and their near-field geometric context. Second, we propose a Scalable Flow-based Interaction Policy for partner-conditioned residual generation. It factorizes dual-humanoid control into a reusable motor foundation and interaction-specific adaptations: partner-aware experts learn corrections over a frozen whole-body motion prior, while conditional flow matching unifies them in a shared residual action space. This enables scalable skill consolidation while retaining pretrained motor capabilities. Extensive experiments demonstrate that OmniInteract improves retargeting fidelity and dual-humanoid coordination while better retaining interaction performance as skill diversity grows. Real-world deployment further demonstrates successful contact establishment, stable whole-body coordination, and recovery from external perturbations.

#26 h=8

Humanoid SRL: Augmenting Humanoids with Supernumerary Limbs for Coordinated Loco-Manipulation

2026-10-19 cs.RO 👤 Zhao-Yuan Gu (h=8)

Fukang Liu, Feiyang Wu, Paul Barsa, Jaehwi Jang, Akshat Jain

Core Contributions

Adds a supernumerary arm to a humanoid to study how morphology changes the feasible action space, enabling manipulation and support at once, such as a third arm while both hands are busy.
Trajectory optimization generates dynamically feasible motions for goal-conditioned RL controllers, trained in simulation and transferred zero-shot to hardware on two configurations.
Diffusion policies trained on hardware demonstrations show vision-guided autonomous execution on selected tasks.
Frames humanoid morphology as an extensible design space rather than a human-like fixed form.

Show abstract

Humanoid robots are envisioned to perform diverse human-level tasks through their anthropomorphic morphology. Their dual-arm configuration enables a broad range of everyday actions. Yet, as in humans, this morphology can constrain loco-manipulation when both arms are occupied or additional contacts are required beyond their reach. Supernumerary robotic limbs (SRLs) offer a way to extend these capabilities, but their integration requires the added limb to be coordinated with whole-body motion and balance. In this study, we develop and integrate a supernumerary arm into a humanoid platform to investigate how morphological augmentation reshapes its feasible action space. Through quantitative analysis and hardware experiments across two configurations, we show how the additional limb expands whole-body interaction capabilities by enabling manipulation and support to be coordinated beyond the constraints of the native dual-arm configuration. To control this augmented morphology, we leverage model-based trajectory optimization (TO) to generate dynamically feasible motions for goal-conditioned reinforcement learning (RL) controllers. The controllers are trained in simulation and transferred to hardware in a zero-shot manner, enabling coordinated whole-body skills for contact-rich tri-arm loco-manipulation. We further show vision-guided autonomous execution on selected tasks using diffusion policies trained from hardware demonstrations. Our findings suggest that humanoid morphology need not be restricted to human-like forms, but can be reimagined as an extensible design space for robotic augmentation.

State Estimation, Mapping & Perception

#1 h=59

Self-Supervised Inertial Odometry

2026-10-19 cs.RO 👤 K. Sreenath (h=59)

Yuheng Qiu, Yutian Chen, Carlton Cort, Weison Huang, Shibo Zhao

Core Contributions

Trains inertial odometry with no ground-truth trajectories: a myopic teacher just integrates the IMU over each short window, and a student network learns from those short-window velocity changes.
Surprisingly, the student predicts velocity more accurately than the integration teacher that supervised it; pose-graph distillation then adds sequence-consistent targets.
Across six datasets (pedestrians, drones, quadrupeds, a humanoid) the same recipe lands within 0.07 m/s average velocity error of a same-architecture supervised baseline.
Why it matters: onboarding a new platform needs only IMU calibration and ordinary operation logs, removing the mocap/GPS rigs that normally gate learned odometry.

Show abstract

Learning inertial odometry for a new platform usually requires ground-truth trajectories. We show that an inertial odometry model can instead be trained from scratch by a myopic teacher that integrates the IMU over each short window independently, a scheme we call integration self-supervision.Interestingly, although supervised only by short-window velocity changes, the student learns to directly predict velocity, and does so more accurately than its integration teacher. Pose graph distillation then refines the model with sequence-consistent targets from a pose graph over many windows. Training on a new platform therefore no longer requires collecting ground-truth trajectories, but only calibrating the IMU and recording data during normal operation. Across six datasets spanning pedestrians, drones, quadrupeds, and a humanoid, the same recipe with no velocity or position labels trains models that come within $0.07$~m/s average velocity error of a same-architecture supervised baseline.

#11 h=20

CAOS-SLAM: Class-Agnostic Object-Aware Real-Time vSLAM for Dynamic Environments

2026-10-19 cs.RO 👤 Shahab Heshmati-alamdari (h=20)

Gaia Borghi, João Carlos Virgolino Soares, Angelo Bratta, Shahab Heshmati-Alamdari, Claudio Semini

Core Contributions

Extends ORB-SLAM3 to reject moving objects whether or not their class is known, avoiding the fixed semantic class lists of earlier dynamic SLAM.
Runs panoptic segmentation only on demand, discovers unknown objects by depth clustering of unclassified regions, and between updates propagates masks with detection or Kalman prediction, depth consistency and appearance.
Up to 59% lower localization error with unknown dynamic objects and up to 88% lower in heavily occluded scenes, at a median 30 FPS on a laptop.
Byproduct: object trajectories and velocities, usable for planning in dynamic scenes.

Show abstract

Robust visual Simultaneous Localization and Mapping (vSLAM) in dynamic environments faces challenges due to the presence of moving objects that may introduce erroneous observations, corrupt the map reconstruction and overall degrade localization accuracy. Existing vSLAM approaches often rely on predefined semantic classes or costly pixel-wise classification, limiting their robustness or applicability in real-world scenarios. This paper presents CAOS-SLAM, a dynamic vSLAM framework that extends ORB-SLAM3 with a hybrid perception pipeline to detect, track, and filter both semantically known and previously unseen moving objects in real-time. On-demand panoptic segmentation is performed for object discovery, with candidate unknown objects instantiated through depth-based clustering of unclassified regions. Between panoptic updates, persistent object tracks are updated using lightweight detection when available or temporal mask propagation based on world-frame Kalman prediction, depth consistency, and appearance similarity. The estimated object motion is then used to reject dynamic features from the SLAM pipeline. Experimental results on multiple dynamic RGB-D datasets and real-world experiments demonstrate improved localization accuracy over state-of-the-art dynamic vSLAM approaches, achieving up to 59% lower localization error in scenes characterized by unknown dynamic objects and up to 88% lower error in heavily occluded scenes, while achieving a median processing rate of 30 FPS on a laptop. Furthermore, the tracking module provides estimates of object trajectories and velocities, supporting potential future applications in perception, navigation, and motion-planning tasks in dynamic environments. Code and dataset will be released upon acceptance.

#15 h=17

Learning Wheel Motion Uncertainty for Degeneracy-Aware LiDAR-Inertial Odometry on Quadrupedal Wheel-Legged Robots

2026-10-19 cs.RO 👤 Yun-Fan Ren (h=17)

Hairuo Wei, Yunfan Ren, Renda Yang, Qiangsheng Huang, Zhongyu Li

Core Contributions

For wheel-legged quadrupeds, keeps analytical wheel rolling predictions but learns per-wheel, direction-dependent process covariances (CovNet) from proprioceptive history, so slip and contact loss are down-weighted.
Degeneracy-aware projection suppresses LiDAR constraints along geometrically weak directions such as long corridors.
CovNet is trained through a differentiable filter on simulated velocity supervision and deployed without fine-tuning; on the six most degenerate sequences it cuts position RMSE 66.2% versus FAST-LIO2 and 31.0% versus fixed covariance.
In a severely degenerate corridor, endpoint error is 0.198 m versus 7.531 m for FAST-LIO2.

Show abstract

Wheel kinematics can support odometry for quadrupedal wheel-legged robots when LiDAR geometry provides weak motion constraints. However, slip and contact loss make rolling predictions unreliable, and contact state alone cannot determine their reliability across motion directions. We propose a tightly coupled LiDAR-inertial-kinematic odometry framework that retains analytical rolling predictions and learns per-wheel, direction-dependent process covariances from proprioceptive history. The estimator maintains moving wheel-center states, allowing the learned uncertainty to regulate the influence of rolling predictions through forward-kinematic corrections. Degeneracy-aware projection (DAP) complements this mechanism by suppressing LiDAR constraints along geometrically weak directions. The learned covariance network (CovNet) is trained through a differentiable filter using simulated velocity supervision and deployed on the real robot without fine-tuning. On the six room sequences with the highest degeneracy ratios in our indoor dataset, the method reduces mean per-sequence position root mean square error (RMSE) by 66.2% relative to FAST-LIO2 and by 31.0% relative to a fixed-covariance variant. In a severely degenerate corridor run, it achieves an endpoint position error of 0.198 m, compared with 7.531 m for FAST-LIO2. These results demonstrate robust odometry under challenging geometric and wheel-motion conditions. We plan to publicly release our code and multimodal dataset. Project website: https://weihairuo.github.io/wheel-motion-uncertainty/

#30 h=7

SpaRV: Sparse Radiance Voxels for Scalable Geometrically Consistent LiDAR-Visual Mapping and Rendering

2026-10-19 cs.ROcs.CV 👤 Zu-Hao Zou (h=7)

Jianheng Liu, Yunfei Wan, Fangming Cheng, Xiangyu Hong, Zuhao Zou

Core Contributions

Supervises a sparse radiance voxel map with irregular LiDAR scans alongside images: LiDAR depth and occupancy anchor geometry where robot trajectories give few viewpoints, with normal and smoothness regularizers between points.
Alternates map training with photometric optimization of camera poses and focal lengths, correcting calibration errors that blur radiance fields.
Bounds GPU memory by keeping one submap at full resolution and the rest at coarser octree levels.
On a single 24 GB GPU it reconstructs a university campus from 63,582 images and simulates other cameras and LiDARs; code at hku-mars/SpaRV.

Show abstract

Robotic simulation, navigation, and inspection need photorealistic and geometrically accurate maps. Radiance fields achieve photorealism, but building them from robot data raises three challenges. A robot's trajectory serves its task, not reconstruction, so images from few viewpoints leave the geometry under-constrained. Residual errors in camera poses and intrinsics blur the map. The map's GPU memory grows with scene size, which limits its level of detail in large scenes. We present \textbf{SpaRV}, a \textbf{Spa}rse \textbf{R}adiance \textbf{V}oxel map rendered by differentiable ray marching, which lets irregular LiDAR scans supervise it alongside images. LiDAR depth and occupancy anchor the under-constrained geometry, and normal and smoothness regularizers keep it smooth between LiDAR points. To correct camera errors, SpaRV alternates map training with photometric optimization of the camera poses and focal lengths. To bound GPU memory in training and rendering, SpaRV holds one submap at a time at full resolution and the rest at coarser octree levels. Experiments show that SpaRV keeps the geometry consistent away from the captured trajectory and sharpens textures by optimizing the camera parameters. On one 24\,GB GPU, SpaRV reconstructs a university campus captured in 63{,}582 images and simulates other cameras and LiDARs. Our code will be released at \url{https://github.com/hku-mars/SpaRV}.

Dexterous & Tactile Manipulation

#3 h=41

HandsOn: A Large-Scale Simulation Benchmark for Generalizable Dexterous Manipulation

2026-10-19 cs.RO 👤 Ning Ding (h=41)

Tianwei Luo, Yiming He, Hongjin Wu, Ting Pan, Debin Ren

Core Contributions

Provides a composable simulator for cross-embodiment dexterity: 5 arms, 6 dexterous hands and about 200 rigid, articulated and deformable assets in randomized scenes.
Pairs 18,000 curated human demonstrations with 66 verified tasks, split into Core-50 (learn and execute) and Transfer-16 (asset/relation transfer, skill composition, deformables with no target-task demos).
The headline is how low the numbers are: the best trained multi-task policy reaches 29.1% on Core-50, and GPT-6 driving palm-and-finger control with move_eef gets 5.2% on Core-50 and 1.9% on Transfer-16 over 660 trials.
Stage-level diagnostics localize failures to contact establishment, action coordination and holding achieved goals, giving a concrete research agenda for dexterous generalization.

Show abstract

General-purpose manipulation requires reliable behavior across embodiments, objects, and task demands. We introduce HandsOn, a composable simulation platform and benchmark for cross-embodiment dexterous manipulation. It supports five robotic arms, six dexterous hands, and approximately 200 rigid, articulated, and deformable assets in randomized scenes. The benchmark pairs 18,000 curated human demonstrations with 66 human-verified tasks: Core-50 supports demonstration-based learning and execution tests, while Transfer-16 evaluates asset and relation transfer, skill composition, and deformable manipulation without target-task training demonstrations. Qualitative real-robot rollouts illustrate sim-to-real transfer. Across diffusion, vision-language-action, world-action, and prompted VLM-based methods, the best trained multi-task policy achieves 29.1% Core-50 success. Without benchmark finetuning, GPT-6 using move_eef for palm-and-finger control achieves 5.2% on Core-50 and 1.9% on Transfer-16 over 660 trials. Trained-policy stage diagnostics expose failures in contact establishment, action coordination, and maintaining achieved goals, with additional sequencing errors in composed tasks. HandsOn provides a shared basis for investigating manipulation generalization and the limits of reliable physical interaction.

#7 h=28

DexOAK: Object-Aware Kinematic Retargeting

2026-10-19 cs.RO 👤 Robert K. Katzschmann (h=28)

Federica Spinola, Robert Katzschmann, Cordelia Schmid

Core Contributions

Retargets human hand motion to robot hands by matching the object-relative interaction geometry (a joint hand-object graph Laplacian), not just hand pose, so contacts remain consistent.
Amortizes the optimization into a feed-forward network conditioned on keypoints and object point clouds, then fine-tunes with contact, non-penetration and self-collision losses; no per-frame optimization at test time.
On the Inspire hand, success rises from 39.5% to 48.4% on OakInk-v2 and ADD-AUC from 62.7% to 77.9% on ARCTIC, with zero-shot transfer between the two.
On hardware 46/56 open-loop replays succeed over five tasks, making mocap human data a practical source of dexterous robot references.

Show abstract

We introduce DexOAK, a learning-based object-aware kinematic retargeting framework for generating contact-consistent robot hand references from human motion-capture data. DexOAK first minimizes an interaction-graph Laplacian objective over a joint hand-object graph, so the robot reproduces the demonstrated object-relative interaction geometry rather than the human hand pose alone. These per-frame solutions form seed retargeted references, from which a feed-forward network conditioned on human keypoints and object point clouds learns the retargeting function. Object-aware finetuning then refines the network under contact, non-penetration, and self-collision losses that position the fingers on the demonstrated contacts and reduce artifacts destabilizing physics-based refinement. Once trained, the network transfers to unseen data by inference alone, without per-frame optimization. DexOAK references improve dynamic retargeting over the state-of-the-art across bimanual tasks with rigid and articulated objects. On the Inspire hand, success rises from 39.5% to 48.4% on OakInk-v2, and ADD-AUC rises from 62.7% to 77.9% on ARCTIC. Ablations trace the gains to object-aware kinematic retargeting and the amortized network refined with contact losses. Trained on OakInk-v2, it transfers to ARCTIC zero-shot and nears in-domain performance after fine-tuning. On hardware, 46/56 open-loop replays succeed across five tasks, enabling dexterous robot data generation from human data. Visualizations and interactive results are at: https://fspinola.github.io/dexoak/.

#12 h=18

Robotic Physical Comparison of Biologically Motivated and Counterfactual Sensory Arrangements using Biomimetic Skin

2026-10-19 cs.RO 👤 Kento Kawaharazuka (h=18)

Akihiro Miki, Shun Hasegawa, Yoshimoto Ribayashi, Yuta Sahara, Kento Kawaharazuka

Core Contributions

Builds a layered biomimetic skin with receptor-analog sensors in compliant media so that spatial arrangements of mechanoreceptors can be compared physically.
Tests biologically motivated layouts against counterfactual ones that biology never produced, which is impossible in living tissue.
On weak-contact detection, two-point discrimination and texture identification, layouts reflecting biological organization outperformed the alternatives examined.
Offers a robotics route to questions about why tactile receptors are arranged as they are, and a design guide for artificial skin.

Show abstract

Human skin contains diverse mechanoreceptors, including Merkel cells and Pacinian corpuscles, each localized in specific regions. Although their spatial organization is known to be shaped by molecular signaling and neural induction during development, experimentally examining its functional significance under controlled conditions in living systems remains challenging. In particular, physically comparing alternative receptor arrangements that do not occur biologically is difficult. In this study, we developed a layered biomimetic robotic skin designed to mimic key structural features of human skin tissue by embedding receptor-analog sensors within compliant soft media. Using this platform, we implemented both biologically motivated and counterfactual sensory arrangements and quantitatively evaluated their performance across three passive tactile tasks: detection of weak contacts, a two-point discrimination-like assessment, and texture identification. Arrangements reflecting key features of the biologically motivated sensory organization exhibited higher performance than the alternative configurations examined. This study addresses a biological question concerning how spatial organization influences tactile function by presenting a biomimetic robotic approach that enables physical and empirical comparison of biologically motivated and counterfactual sensory arrangements that are difficult to investigate in living systems.