🤖 Robotics arXiv Digest

Thursday, October 1, 2026

📄 30 papers 📂 7 research areas Generated by Claude

🔭 Research Landscape

Today's strongest current is agents that improve robots without retraining the underlying model, by practicing in digital twins. Recova (#3) and Reconstruct, Practice, Go Real (#9) arrive at nearly the same loop from different directions: reconstruct the scene in simulation, let an LLM agent diagnose failures with privileged state, keep only verified fixes, then go to hardware. Recova distills the fixes into separate task and recovery policies via DAgger (real-robot success 23.8%→87.5%, human interventions 87.5%→0% over four rounds), whereas RPG keeps weights frozen and grows a symbolic skill library plus system prompt (28.6%→95.0% over 15 practice rounds, 30/30 physical trials). The two benchmarks in the batch temper the optimism: Embodied Agent Arena (#23) finds frontier VLMs strongest at metric estimation and contact localisation but weakest at coordinated goal completion, and HumanoidToolBench (#4) shows GR00T N1.7 picks a reasonable tool far more often than it actually finishes the task. The common lesson is that reasoning is no longer the bottleneck — grounded execution and verification are, which is exactly what the twin-based practice loops target.

A second thread is one generative backbone for many robot problems — and making it fast and physically honest. World Motion Models (#1) collapses prediction, infilling, MPC, IK, retargeting and policy learning into masks over a single flow-matching model on SE(3) trajectories; UniWAM (#19) fuses a VLM reasoner, video generator and action head and reports a log-linear scaling law for human-robot co-training. Two papers then attack the cost and reliability of these models: Kinematic MeanFlow (#12) makes one-step action generation work for foundation-model action heads (GR00T-N1.6 action-head latency −67–74%), while PhysicsLENS (#20) shows that video world models usually ignore hidden properties like weight or friction stated in the prompt (34 of 47 plausible-looking videos). Flow matching itself is now a generic tool — it also appears as Riemannian flow matching on meshes (#6) and as a resolution-consistent Jacobian field for a soft finger (#25).

Third, the batch has an unusually rigorous look at sampling-based optimisation. Two independent CEM papers are complementary: In CEM, a World Model Is Also a Proposal Mechanism (#13) shows that a world model's ranking errors corrupt not just the current choice but the next iteration's proposal distribution, while TierCEM (#17) changes elite selection itself to enforce strict constraint priorities without tuned penalty weights. BLT* (#2) brings RRT*-style rewiring into belief space using a Wasserstein metric, and the finite-data safety certificate (#22) asks which commands are provably safe for every model consistent with limited data. On the coordination side, decentralisation is the default: SLM-powered UAV swarms (#5) find that letting the language model decide what to forward breaks the mission entirely, so dissemination must be deterministic, and Watch, Infer, Coordinate (#14) infers a degraded partner's limits from the team's joint behaviour rather than the partner alone. Field and medical systems (#7, #8, #10, #18) round out the day with long-duration and in-vivo-adjacent validation that is rare in learning-centric papers.

Generative World & Action Models

Unified generative backbones over motion, video and action — and the latency and physics-fidelity problems they bring.

#1 World Motion Models: Flexible Sequence Modeling of SE(3) T...
#12 Kinematic MeanFlow: One-Step Action Generation Policy for...
#16 Continuous Conditioning of VLAs with Augmenting EMG and Vi...
#19 UniWAM: Unified World-Action Model
#20 PhysicsLENS: Diagnosing Physical Property Blindness in Vid...

Agentic Self-Improvement & Embodied Benchmarks

LLM/VLM agents that practise in digital twins, and benchmarks measuring how far frontier models are from robot generalists.

#3 Recova: Agent-Guided Failure Recovery for Autonomous Robot...
#4 HumanoidToolBench: Benchmarking Humanoid Tool Use from Sel...
#9 Reconstruct, Practice, Go Real: Guided Self-Improvement fo...
#23 Are Frontier VLM Agents Ready to Be Robot Generalists? An...

Imitation, RL & Motion Representation

Learning from observation, new RL objectives, movement primitives and retargeting for dexterous and surface-bound skills.

#6 Robot Learning on Discrete Surfaces: Theory and Applicatio...
#11 Experience-Based Feasibility-Aware Generative Adversarial...
#24 FERPO: Forward Entropy-Regularized Policy Optimization
#25 Learning a Resolution-Consistent Jacobian Field for Bio-In...
#27 Interaction-Stiffness-Guided Basis Allocation in Dynamic M...
#30 FlashDexRetarget: Accelerating Dexterous Manipulation Data...

Planning, Optimisation & Certified Safety

Belief-space planning, a close look at the cross-entropy method, and data-driven safety certificates.

#2 BLT*: Informed Belief Localization Trees for Uncertainty-A...
#13 In CEM, a World Model Is Also a Proposal Mechanism
#17 Cross-entropy optimization with prioritized constraints
#22 Finite-Data Safety Informativity Under Dynamic Asymmetric...

Multi-Robot & Multi-Agent Coordination

Decentralised swarms, partner-constraint inference, multi-humanoid whole-body control and competitive exploration.

#5 Managing Context and Communication in Distributed Agentic...
#14 Watch, Infer, Coordinate: Inferring Robot Partner Constrai...
#21 MASkillBlender: Decentralized Whole-Body Coordination for...
#26 OpenSpace Lab Solution to the IROS 2026 Indoor Exploration...

Simulation, Digital Twins & 3D Assets

Simulators and datasets that add missing physics — hydrodynamics, metric 3D, tactile micro-geometry and heat.

#15 H-SPAR: Hydrodynamic-aware Simulation for Particle Transpo...
#28 3DROID: A Renderable 3D Gaussian Dataset with Measured Per...
#29 TouchTherm: Building Multimodal Digital Twins of Objects f...

Field, Medical & Human-Assistive Robotics

Underwater diver guidance, spine surgery, long-term forest monitoring and autonomous colonoscopy.

#7 Towards Physical Underwater Robotic Assistance for Scuba D...
#8 Ex vivo breach detection using electrical conductivity dur...
#10 ALFRED: Requirement-driven development of an open-source m...
#18 ColoACT: Multi-Cue Action Chunking for Smooth Autonomous C...

Generative World & Action Models

h=62

World Motion Models: Flexible Sequence Modeling of SE(3) Trajectories

📅 2026-10-01cs.ROcs.CV👤 Angjoo Kanazawa (h=62)

Jiahui Lei, Qianqian Wang, Trevor Darrell, Angjoo Kanazawa

Core Contributions

Argues that a set of rigid SE(3) trajectories is a sufficient common currency for 4D scenes — articulated objects, bodies, hands, cameras and robot actions all live in one token space instead of task-specific representations.
Per-token noise levels in flow matching turn the model into an any-to-any conditional generator: forecasting, infilling, MPC, IK, cross-embodiment retargeting and policy learning become different masks on one network rather than six separately trained models.
A context-token mechanism handles non-sequential conditioning, so arbitrary numbers of entities and time steps can be mixed — the property that makes cross-embodiment transfer possible.
Matters because it reframes a 'world model' as a sparse pose prior rather than a pixel generator, which is far cheaper and directly actionable for control; validated across 6 vision and robotics applications.

Show abstract

Equipping artificial agents with spatial intelligence requires a comprehensive generative prior over the dynamic 3D world. We propose World Motion Models (WMMs) that capture "what was, is, and will be where across time" via sparse SE(3) pose trajectories. WMMs are built on the observation that elements of dynamic scenes can be well approximated by a set of rigid SE(3) trajectories, a minimal yet expressive primitive for 4D modeling. This representation unifies articulated objects, human bodies, hand-object interactions, piecewise-rigid scene dynamics, camera motion, and even robot states and actions into a single shared space. Given this representation, we cast the joint distribution of these entities as a flexible sequence modeling problem, utilizing flow-matching with per-token noise levels. Coupled with a context token mechanism for non-sequential conditioning, this formulation supports any-to-any marginal conditioning across an arbitrary number of entities and time steps. Tasks such as future prediction, motion infilling, model-predictive control, inverse kinematics, cross-embodiment retargeting, and policy learning all reduce to the application of different masks over the same network. Experiments on 6 diverse applications of 3D vision and robotics demonstrate the versatility and flexibility of WMMs with strong performance.

#12

h=25

Kinematic MeanFlow: One-Step Action Generation Policy for Robotic Foundation Models

📅 2026-10-01cs.ROcs.AIcs.CLcs.CV👤 An-Bang Yao (h=25)

Jiawei Fan, Sifeng Wang, Yuqing Hou, Anbang Yao

Core Contributions

Diagnoses why MeanFlow collapses in robotic foundation models: the velocity field's local acceleration spikes near the end of denoising and its magnitude spread widens across samples.
Uses a kinematic identity to split MeanFlow's time-derivative term at an intermediate point, so early- and late-stage dynamics are modelled separately and error does not compound.
Achieves one-step action generation both from scratch and when fine-tuning, while matching or beating multi-step flow matching in most settings.
Concrete latency gains on GR00T-N1.6: action-head latency −67.5% to −74.4% on L40 and Jetson Orin, end-to-end −30.3% to −54.9% — important for edge deployment.

Show abstract

In this paper, we study how to achieve one-step action generation in Robotic Foundation Models (RFMs), aiming to overcome the high inference latency of multi-step flow matching. MeanFlow provides a promising framework for this goal, yet its direct application leads to performance collapse. We discover that this stems from two distinctive dynamics exhibited in the RFM velocity field: (1) the ``local acceleration" exhibits stability early on, but surges sharply towards the end of the denoising process, and (2) the spread of its magnitudes across samples widens as denoising progresses. To address these issues, we introduce Kinematic MeanFlow (K-MF), a novel one-step action policy tailored for RFMs. Specifically, grounded in a kinematic identity, K-MF decouples the time derivative term in the MeanFlow formulation into two sub-interval terms separated by an intermediate point. This decoupled formulation enables the two terms to capture early-stage and late-stage denoising dynamics, respectively, while mitigating the error amplification across the process. As a result, our K-MF empowers RFMs to achieve one-step action generation in both training from scratch and fine-tuning paradigms across diverse tasks, while outperforming multi-step flow matching in most settings. In terms of inference efficiency, K-MF reduces action-head latency of GR00T-N1.6 by 67.5%~74.4% across L40 and Jetson Orin in eager and compiled modes, yielding end-to-end latency reductions of 30.3%~54.9%. Code will be available at https://github.com/IntelChina-AI/K-MF.

#16

h=23

Continuous Conditioning of VLAs with Augmenting EMG and Visual Task Descriptors

📅 2026-10-01cs.CVcs.RO👤 M. Fifer (h=23)

Edward W. Staley, Connor O. Pyles, Rahul Hingorani, Frank Camargo, Griffin Milsap

Core Contributions

Challenges the assumption that language is the right channel for task specification in VLAs, especially in cluttered or ambiguous scenes.
EC-VLA feeds 8-channel EMG envelopes as continuous conditioning alongside proprioception; VA-VLA adds visual segmentation annotations to the images.
On a cube-selection task with three participants, both match language-prompted baselines in-distribution and substantially outperform them in cluttered OOD scenes.
Points toward assistive and prosthetic settings where muscle signals, not words, are the natural interface to a VLA.

Show abstract

Vision-Language-Action (VLA) models rely strongly on language for describing task information, despite having multimodal inputs. We hypothesize that other modalities in the state space may present opportunities for supplemental task conditioning, which may be particularly relevant in cluttered or otherwise ambiguous scenes. We introduce two tuned models to test this hypothesis: (1) an electrophysiology-conditioned VLA (EC-VLA) that incorporates 8-channel electromyography envelopes as continuous conditioning input concatenated to the proprioceptive vector, and (2) a visually-annotated VLA (VA-VLA) that incorporates visual segmentation annotations to the image inputs. On a cube-selection task evaluated across three participants, EC-VLA matches a language-prompted baseline in uncluttered, in-distribution conditions and substantially outperforms it in cluttered, out-of-distribution scenes. Similarly, VA-VLA shows modest improvements over a language-prompted baseline in in-distribution scenes with substantial improvement in cluttered, out-of-distribution trials. Together, these results provide strong evidence for the potential benefit of task-conditioning beyond language.

#19

h=21

UniWAM: Unified World-Action Model

📅 2026-10-01cs.RO👤 Peng-Wei Wang (h=21)

Jiayi Chen, Wenxuan Song, Jingbo Wang, Shuai Zhou, Xicheng Gong

Core Contributions

Combines a physical reasoner (VLM), a world generator (video) and an action predictor so a single model gets both semantic reasoning and spatiotemporal dynamics priors.
Expresses low-level actions in natural language and routes VQA, human egocentric and robot data to the components each is best for, preserving the VLM's language ability.
Future-visual-noise augmentation and history-conditioned flow matching cut denoising steps by reducing reliance on precise future prediction.
Reports a log-linear scaling law for unified human-robot co-training — quantitative evidence that cheap human video keeps paying off at scale.

Show abstract

Vision-language-action models benefit from the understanding and reasoning capabilities of pretrained vision-language models, but action-only supervision provides limited grounding in world dynamics. Conversely, world-action models inherit spatiotemporal priors from video generation models, yet remain limited in semantic understanding and reasoning under distribution shifts. We introduce UniWAM, a unified architecture that integrates a physical reasoner, a world generator, and an action predictor to jointly learn semantic understanding of the physical world, visual generation, and action prediction. To ensure the quality of the training data, we developed a rigorous data cleaning and annotation pipeline for both human egocentric data and robot data. To adapt the vision-language component to embodied tasks while preserving its inherited language capabilities, we represent low-level actions in natural language and introduce a pre-training recipe that assigns complementary supervision from visual question answering (VQA) data, human egocentric data, and robot demonstrations to the appropriate model components. During post-training, future visual noise augmentation reduces reliance on precise future predictions, while history-conditioned flow matching uses encoded action history to initialize action generation. Together, these designs significantly reduce denoising steps while maintaining performance. UniWAM achieves state-of-the-art (SOTA) performance across multiple evaluations, including in-distribution performance, robustness, generalization, instruction following, and long-horizon task execution. Furthermore, we uncover a log-linear scaling law of unified human-robot co-training, demonstrating the effectiveness of large-scale pre-training on a mixture of human and robot data.

#20

h=21

PhysicsLENS: Diagnosing Physical Property Blindness in Video Generation Models

📅 2026-10-01cs.CVcs.RO👤 Ransalu Senanayake (h=21)

Isaiah Milkey, Som Sagar, Aditya Taparia, Xinyuan Liu, Jiqing Wen

Core Contributions

Highlights a blind spot: video benchmarks test whether physics looks right, not whether hidden properties such as weight, viscosity or friction are respected.
Uses matched scenario pairs with the same first frame and task but different stated physics, drawn from public robot videos across seven domains (collision, gravity, momentum, friction, deformation, fluid, causality).
Over 400 human labels across four video generators: 34 of 47 plausible-looking videos ignore the stated property, and stating it barely changes plausibility.
A warning for anyone using video world models for policy evaluation or planning — visual realism is not physical correctness.

Show abstract

Reliable video world models could provide scalable predictive environments for robot learning, planning, and evaluation. However, generated robot videos can violate physical principles and complete tasks through physically implausible behavior, limiting their reliability for robot learning and planning. Current video-generation benchmarks exclude physics that are inherently hidden by visuals (e.g., weight, viscosity, friction). Due to this, video models are evaluated on the fidelity of physics, not the underlying accuracy of physics. We introduce PhysicsLENS, a dataset and benchmark for evaluating plausibility of physical properties grounded in robotics. PhysicsLENS uses matched scenario pairs that hold the same conditioning frame and task, while varying underlying physics in the scene description. Scenarios are curated from public robot video sources and annotated across seven physical domains: collision, gravity, momentum, friction, deformation, fluid, and causality. We evaluate across four video generation models, producing over 400 human-annotated labels. Results show that plausible-looking videos often ignore the stated property (34 of 47), and that stating the property lowers plausibility only slightly and not significantly.

Agentic Self-Improvement & Embodied Benchmarks

h=43

Recova: Agent-Guided Failure Recovery for Autonomous Robotic Manipulation

📅 2026-10-01cs.RO👤 Yu-Ke Zhu (h=43)

Isabella Liu, An-Chieh Cheng, Johan Bjorck, Zhiding Yu, Hongxu Yin

Core Contributions

Treats recovery from failure as its own learnable skill: an agent in a reconstructed digital twin diagnoses failures, writes and tests corrective programs, and harvests successful task and recovery rollouts for separate policies.
At deployment it monitors progress, triggers a learned or programmatic recovery, verifies the scene is restored, then resumes — falling back to a human demo only when nothing fits, which then enters the DAgger loop.
Simulation: 78.8% (LIBERO-Pro) and 64.9% (MolmoSpaces) mean success vs. 71.7% and 38.0% for the strongest baselines — the gap is largest where scenes are most diverse.
Real world across four parallel workstations: DAgger lifts success 23.8%→77.5%, recovery skills add another 10 points to 87.5%, and human intervention on one task drops from 87.5% to 0% in four rounds — failures become reusable capabilities.

Show abstract

Manipulation failures can leave scenes in states from which a task policy cannot recover. Learning corrective behaviors requires scalable failure exploration and physical grounding. We present Recova, an agent-guided framework that jointly develops task execution and recovery in a reconstructed digital twin, then verifies and refines both through real-world experience. In the twin, the agent diagnoses failures, tests corrective programs, and collects successful task and recovery rollouts for separate policies. During deployment, it monitors progress, invokes a learned or programmatic recovery, verifies scene restoration, and resumes execution. When no suitable recovery is available, a human demonstration resolves the failure and enters the learning loop, allowing the system to expand its recovery capabilities. Physical rollouts and human demonstrations are routed to the corresponding policy for DAgger training. Across six LIBERO-Pro settings and four MolmoSpaces categories, Recova achieves 78.8% and 64.9% mean success, compared with 71.7% and 38.0% for the strongest baselines. With parallel collection across four real-robot workstations, DAgger fine-tuning raises mean success from 23.8% to 77.5%, and recovery skills further raise it to 87.5%. Over four collection rounds on one task, observed human intervention falls from 87.5% to 0%. Together, these results show how agent-guided recovery turns failures into reusable capabilities, improving robustness while progressively reducing human intervention. Project page: https://www.liuisabella.com/Recova

h=38

HumanoidToolBench: Benchmarking Humanoid Tool Use from Selection to Mobile Execution

📅 2026-10-01cs.ROcs.AI👤 Idan Szpektor (h=38)

Kyochul Jang, Seohyeon Park, Ohchul Kwon, Sangjun Park, Junhyeok Choi

Core Contributions

First benchmark that jointly tests tool selection, manipulation and locomotion on a humanoid: 18 tasks × 3 scenarios × 3 execution levels × 2 tool-set modes.
Releases ToolBook, 3.1k demonstrations from both simulation and a real Unitree G1, so policies can be trained and compared on the same tasks.
Evaluating seven policies in sim and three on hardware exposes a large gap between choosing an appropriate tool and completing the task with it.
Probes of GR00T N1.7 show selection accuracy drops on unseen tools and the policy keeps executing even under unrelated instructions — evidence of weak language grounding in current humanoid foundation models.

Show abstract

As robotic hardware and learning methods advance, humanoids need tools to perform tasks beyond their inherent physical limits. Successful tool use requires selecting a suitable tool and coordinating manipulation and, when needed, locomotion to complete the task. Existing benchmarks do not jointly evaluate these capabilities on a humanoid. We introduce HumanoidToolBench, an 18-task benchmark spanning three scenarios, three execution levels, and two tool-set modes, together with ToolBook, a dataset of 3.1k demonstrations collected in simulation and on a real Unitree G1. Evaluation of seven policies in simulation and three on the real robot reveals substantial gaps between selecting a suitable tool and completing the task. Focused GR00T N1.7 probes show reduced selection accuracy on unseen tools and continued task execution under unrelated instructions. Code and data are available at https://snu-pi.github.io/HumanoidToolBench/.

h=26

Reconstruct, Practice, Go Real: Guided Self-Improvement for Embodied Agents

📅 2026-10-01cs.ROcs.AIeess.SY👤 Nika Haghtalab (h=26)

Yen-Jen Wang, Haozhe Jiang, Shuying Deng, Haoru Xue, Weirui Ye

Core Contributions

RPG improves a robot system purely at the prompt-and-skill level: it mines capabilities from an offline dataset, builds matching practice tasks in simulation, and diagnoses failures using execution feedback, privileged sim state and dataset videos.
New symbolic skills and prompt revisions are only kept after cross-task evaluation, preventing a fix for one task from silently breaking others — a key difference from naive self-refinement.
On held-out initialisations of 22 manipulation tasks, success climbs from 28.6% to 95.0% over 15 rounds, ahead of ASPIRE (75.5%) and a CaP-Agent0 baseline (60.0%).
After a shared calibration step the frozen system succeeds in 30/30 physical trials — evidence that practice in reconstruction can transfer without any weight updates.

Show abstract

Building reliable robot capabilities across diverse tasks requires substantial human effort to develop and maintain skills, design rewards, and integrate perception with control. We present Reconstruct, Practice, Go Real (RPG), a framework for autonomous improvement of robot execution systems without updating model weights. RPG identifies manipulation capabilities in an offline dataset and constructs related practice tasks in simulation. During practice, RPG uses execution feedback, privileged simulator state, and available dataset videos to diagnose failures. It develops new reusable symbolic skills, refines existing skills, and revises the system prompt based on these diagnoses. Cross-task evaluation tests individual candidate changes and merged revisions before they are retained for reuse. At test time, a multimodal LLM uses the resulting system prompt and skill library to coordinate perception and robot control. On held-out initializations of 22 manipulation tasks, RPG improves task success from 28.6% after the first practice round to 95.0% after 15 rounds, outperforming all evaluated baselines, including ASPIRE (75.5%) and CaP-Agent0 powered by GPT-6 Astra Pro (60.0%). After a common calibration and hardware-adaptation procedure, the frozen system succeeds in all 30 physical trials, with ten trials on each of three tasks. Project Website: https://rpg-robot.github.io/

#23

h=20

Are Frontier VLM Agents Ready to Be Robot Generalists? An Empirical Study with the Embodied Agent Arena

📅 2026-10-01cs.RO👤 Yin-Chuan Li (h=20)

Haojian Huang, Pukun Zhao, Zexi Li, Yehang Zhang, Yangkai Wei

Core Contributions

Embodied Agent Arena collects 1,000 cases from 32 sources plus a new geometry benchmark (GeoProbe) to test where local skills do or do not translate to task success.
A minimal harness separates metric precision, functional grounding and native goal completion, so failures can be localised to a specific ability.
Across seven VLMs, the best model's edge is largest in precise estimation and usable-contact localisation.
Coordinated, goal-directed action remains the main gap — useful guidance for where VLM-as-robot-brain research should focus.

Show abstract

Frontier vision-language models (VLMs) combine scene estimation, interaction grounding, and executable actions. Understanding how these abilities support complete robotic tasks is central to evaluating their readiness as robot generalists. We introduce Embodied Agent Arena to examine where local competence supports, or falls short of, complete task success across Geometry, Spatial Reasoning, Affordance, Task Planning, and Manipulation. The arena contains 1,000 cases drawn from 32 established sources and GeoProbe, our new benchmark for geometric estimation on Blender renders and real-scene images. A minimal harness preserves source observations and operations while separating metric precision, functional grounding, and native goal completion. We evaluate seven VLMs, analyze Astra's task-specific advantages, and compare richer-observation execution protocols and multi-round review. Across the arena, Astra's advantage is strongest in precise estimation and usable-contact localization; completing coordinated, goal-directed actions remains the key gap to robot generalism.

Imitation, RL & Motion Representation

h=32

Robot Learning on Discrete Surfaces: Theory and Applications

📅 2026-10-01cs.RO👤 D. Fontanelli (h=32)

Matteo Dalle Vedove, Fares J. Abu-Dakka, Luigi Palopoli, Daniele Fontanelli, Matteo Saveriano

Core Contributions

Builds a discrete Riemannian toolkit — log/exp maps, parallel transport and ambient projections — that stays well-defined on faces, edges and vertices of polyhedral meshes, the native output of CAD and 3D reconstruction.
Rather than treating a surface as a constraint, learning happens intrinsically on it; the same toolkit powers three paradigms: DMPs, Gaussian processes (geodesic kernel) and Riemannian flow matching.
Improved exponential-map computation and parallel-transported forcing terms give DMPs better cross-surface generalisation; mesh-native operators beat spectral baselines for RFM while training faster.
Demonstrated on real robots transferring user-drawn trajectories between surfaces and planning polishing paths on RGB-D-reconstructed meshes — directly relevant to finishing and inspection tasks.

Show abstract

All the objects composing our world are enclosed within surfaces. Yet, most robot learning and motion generation frameworks treat surfaces as constraints ignoring their intrinsic geometry. This gap is acute for polyhedral meshes--the standard output of CAD and 3D reconstruction--whose discrete geometric structure remains unexploited. In this paper, we propose a unified discrete Riemannian framework that enables robot learning directly on polyhedral surface meshes. Using discrete differential geometry, we define logarithmic and exponential maps, parallel transport, and ambient-space projections that remain well-defined across faces, edges, and vertices. We instantiate the framework in three learning paradigms: (i) Dynamic Movement Primitives (DMPs), an improved exponential-map computation and a fixed-tangent-cone forcing-term encoding with parallel transport yield better cross-surface generalisation and stability over prior mesh-based approaches. (ii) Gaussian Process (GP), a geodesic-based kernel with practical admissibility control, enables regression at arbitrary mesh locations without smoothness assumptions. (iii) Riemannian Flow Matching (RFM), mesh-native operators improve generative quality over spectral baselines while reducing training time. The framework is validated in simulation against state-of-the-art methods and demonstrated on two real-robot scenarios: generalising user-drawn trajectories across different surfaces and planning polishing motions on RGB-D-reconstructed surfaces.

#11

h=26

Experience-Based Feasibility-Aware Generative Adversarial Imitation from Observation under Embodiment Mismatch

📅 2026-10-01cs.RO👤 Takamitsu Matsubara (h=26)

Yoshiki Takebayashi, Giovanni Perantoni, Hikaru Sasaki, Matteo Saveriano, Takamitsu Matsubara

Core Contributions

Tackles imitation from state-only human demos when some human motions are simply infeasible for the robot's body.
Unlike methods that need an explicit dynamics model or large exploration datasets to judge feasibility, EF-GAIfO estimates it from the robot's own accumulated experience.
Feasibility is not fixed: as the policy improves and covers more transitions, more demonstrations are admitted — a curriculum that emerges automatically from learning progress.
Validated on simulated locomotion and a real quadruped doing reach-and-grasp, relevant to scaling robot-free demonstration interfaces.

Show abstract

With the increasing use of robot-free demonstration interfaces that provide state trajectories without action labels, imitation from observation has become a promising approach for learning robot behaviors from human demonstrations. However, due to differences in embodiment and dynamics between humans and robots, demonstrated human motions may not be feasible for the robot, potentially degrading policy performance. In this study, we propose Experience-Based Feasibility-Aware Generative Adversarial Imitation from Observation (EF-GAIfO), which estimates the feasibility of state-only demonstrations from the robot's own experience rather than relying on explicit dynamics models or large prior exploration datasets. A key feature of EF-GAIfO is that the notion of feasibility evolves with policy learning: as the policy improves and the robot experiences a broader range of state transitions, the feasible region is progressively expanded, allowing additional demonstrations to be incorporated into learning. This enables feasibility-aware imitation that adapts to the current stage of policy learning, rather than relying on a pre-designed feasibility criterion. We validate the effectiveness of EF-GAIfO on a locomotion task in simulation and on a real quadruped robot performing a object-reaching-and-grasping task.

#24

h=19

FERPO: Forward Entropy-Regularized Policy Optimization

📅 2026-10-01cs.LGcs.AIcs.ROstat.ML👤 Majid Khadiv (h=19)

Sebastian Sanokowski, Alireza Sarmadi, Majid Khadiv

Core Contributions

Avoids differentiating a learned critic with respect to actions, since accurate value predictions do not guarantee accurate action gradients.
Derives an entropy- and KL-regularised target action distribution and fits the actor with a forward-KL objective estimated by self-normalised importance sampling over rollout actions.
Forward KL covers multiple high-value modes (unlike mode-seeking reverse KL), improving exploration; the KL term keeps importance weights stable.
Competitive with sample-efficiency gains on MuJoCo Playground and ManiSkill, with faster actor updates than REPPO.

Show abstract

Several state-of-the-art methods for online reinforcement learning in continuous control improve policies using action gradients of a learned critic. However, critics are typically trained to predict returns, and accurate value predictions do not necessarily yield accurate action derivatives, potentially leading to unreliable policy updates. We propose Forward Entropy-Regularized Policy Optimization (FERPO), an on-policy maximum entropy reinforcement learning algorithm that performs policy improvement using critic values without differentiating the critic with respect to actions. FERPO derives an optimal target action distribution from a policy-improvement objective regularized by entropy and Kullback-Leibler (KL) divergence. We then fit the actor to this target by minimizing a forward-KL objective, estimated using self-normalized importance sampling (SNIS) with actions drawn from the rollout policy. By limiting the target distribution's deviation from the rollout policy, the KL regularization helps keep these importance weights well behaved. In contrast to reverse-KL objectives, which can favor a subset of the target distribution's modes, the forward-KL objective encourages coverage of multiple high-value modes and thereby promotes exploration. Experiments and ablations on MuJoCo Playground and ManiSkill show competitive performance and sample-efficiency gains. Computational benchmarks also demonstrate faster actor updates than Relative Entropy Pathwise Policy Optimization (REPPO).

#25

h=17

Learning a Resolution-Consistent Jacobian Field for Bio-Inspired Rigid-Soft Finger

📅 2026-10-01cs.RO👤 Zi-Rong Luo (h=17)

Tianyou Liang, Haisen Zeng, Shanjun Chen, YiMing Zhu, Zhongyue Lu

Core Contributions

Point-wise Jacobian estimates for tendon-driven rigid-soft fingers depend on sensor and controller sampling rates; JFM learns a resolution-consistent Jacobian field instead.
Models actuation-to-motion as a conditional flow, supporting both single-step prediction and continuous ODE rollout at any temporal resolution.
Cuts global average single-step RMSE by over 53% vs. discrete Jacobian learning and suppresses outlier errors.
Under sparse sampling (stride 8), long-horizon rollout reduces median RMSE by 14.43% and error variance by 24.87% — useful for offline trajectory optimisation.

Show abstract

Bio-inspired tendon-driven rigid-soft coupled dexterous fingers exhibit strong nonlinearity and configuration-dependent sensitivity, making accurate modeling challenging. In discrete-time control, Jacobian-based kinematic algorithms typically rely on point-wise local linear approximations, which makes their performance sensitive to sensor sampling frequency and controller update frequency. To address this issue, we propose Jacobian Flow Matching (JFM), a structured learning framework based on Conditional Flow Matching (CFM), to learn a resolution-consistent Jacobian field that models actuation-to-motion transitions as a dynamical flow. The proposed framework supports both single-step prediction and continuous rollout via ODE integration, enabling consistent inference across temporal resolutions. Experiments on a tendon-driven rigid-soft finger show that the proposed method suppresses outlier errors and improves single-step prediction accuracy, reducing the global average RMSE by over 53% compared with a baseline discrete Jacobian learning approach. For long-horizon prediction, trajectories recovered via ODE integration achieve higher fidelity under sparse sampling (Stride = 8), reducing the RMSE median by 14.43% and the error variance by 24.87%. These results demonstrate that the learned flow-based Jacobian field provides an effective local model for offline multi-step trajectory optimization in rigid-soft coupled nonlinear systems.

#27

h=17

Interaction-Stiffness-Guided Basis Allocation in Dynamic Movement Primitives for Efficient Skill Transfer

📅 2026-10-01cs.RO👤 Si-Lu Chen (h=17)

Chan Xu, Silu Chen, Dehao Wang, Xiyu Chen, Dexin Jiang

Core Contributions

Classic DMPs spread basis functions uniformly in time, wasting capacity on easy segments and under-fitting precision-critical ones.
SC-DMP builds a stage-criticality index from operator-robot interaction stiffness and cross-demonstration task-space variability.
Redistributes basis centres via inverse cumulative criticality and refines bandwidths, giving dense representation only where the task demands it.
Beats DMPs, ProMPs, ProDMP, GP-MP and KMP on reproduction, endpoint generalisation and task-critical accuracy across handwriting and three real-robot tasks while keeping DMP stability.

Show abstract

Dynamic Movement Primitives (DMPs) provide a compact and stable formulation for trajectory representation and generalization in robot skill learning. However, their predefined basis layout limits the allocation of approximation capacity according to stage-dependent precision requirements. To address this issue, this article proposes Stage-Criticality-Guided Dynamic Movement Primitives (SC-DMPs) with adaptive basis allocation for precision-critical skill learning. Operator-robot interaction stiffness and a trajectory-consistency cue derived from cross-demonstration task-space variability are integrated to construct a stage-criticality index. Guided by this index, basis centers are redistributed in normalized time through inverse cumulative criticality and mapped to the canonical phase domain, while their bandwidths are refined to adjust local approximation support. This enables denser and more flexible representation at high-criticality stages while retaining sparser allocation elsewhere. Experiments on handwriting trajectories and three real-robot tasks show that the inferred criticality is concentrated in geometrically demanding and task-constrained regions. Comparisons with DMPs, ProMPs, ProDMP, GP-MP, and KMP demonstrate improved trajectory reproduction, endpoint generalization, and task-critical accuracy while retaining a compact model and the stable structure of classical DMPs.

#30

h=14

FlashDexRetarget: Accelerating Dexterous Manipulation Data Generation through Multi-Motion Retargeting

📅 2026-10-01cs.RO👤 Jaegul Choo (h=14)

Kyungmin Lee, Sibeen Kim, Dongyoon Hwang, Yoonsang Oh, Donghu Kim

Core Contributions

Turns human hand-object demos into physically feasible robot motions with an RL retargeter designed for both high success and low training cost.
Combines object point clouds, hand-object distance features and future-trajectory encodings with rewards on both object motion and hand-object relations.
Separate left/right-hand actor-critics and an adapted FlashSAC give 90% success on 50 motions — ~2.5× sampling-based baselines — with up to 100× less compute than RL baselines.
Gains hold on XHand and Sharpa Wave Hand and scale to 1,000 motions, making large dexterous datasets cheaper to produce.

Show abstract

Human hand-object demonstrations offer a reusable source of dexterous robot manipulation data, but transferring them across embodiments requires physically feasible retargeting. Existing physics-based approaches face limitations in retargeting success, motion-specific training efficiency, or both. To address these limitations, we introduce FlashDexRetarget, an RL-based framework for high-success, efficient dexterous motion retargeting. To make the demonstrated interaction easier to learn, we combine object point-cloud observations, hand-object distance features, and future trajectory encodings with complementary rewards that supervise object motion and reference hand-object relationships. To further accelerate learning, we employ separate left- and right-hand actor critic networks and adapt the off-policy algorithm, FlashSAC to dexterous motion tracking. On a benchmark of 50 motions spanning single-object and two-object interactions, FlashDexRetarget achieves a 90% success rate, approximately 2.5x that of the evaluated sampling-based baselines, while requiring up to 100x less training compute than the evaluated RL-based baselines. Evaluations on both XHand and Sharpa Wave Hand show consistent gains, and component-wise ablations examine the contributions of our design choices. Beyond the 50-motion benchmark, experiments with 200, 500, and 1,000 motions demonstrate that our method remains stable at larger scales and produces successful retargeted motions more efficiently as the training set grows. Qualitative replay results using real-world-captured demonstrations further illustrate the applicability of our framework to recorded human manipulation. Videos and code are available at https://davian-robotics.github.io/FlashDexRetarget/

Planning, Optimisation & Certified Safety

h=49

BLT*: Informed Belief Localization Trees for Uncertainty-Aware Planning on Digital Twins

📅 2026-10-01cs.RO👤 T. Barfoot (h=49)

Elliot Preston-Krebs, Abhishek Goudar, Timothy D. Barfoot

Core Contributions

Lifts RRT*/Informed RRT* into belief space by using the 2-Wasserstein distance between isotropic Gaussian beliefs as the steering/rewiring metric.
Unlike prior belief-space planners that re-propagate observations every time an edge is rewired, BLT* reuses previously computed measurement information — the key to scaling to large outdoor maps.
Handles probabilistic collision constraints and point-cloud localisation directly, and ships a pipeline for semantically labelled digital twins of real environments to plan on.
Finds initial solutions faster in most maps while keeping competitive cost convergence, making uncertainty-aware planning practical at site scale.

Show abstract

We present Informed Belief Localization Trees* (Informed BLT*), a sampling-based belief space planning (BSP) algorithm that scales to large outdoor digital twins with point-cloud observations. We adapt RRT* and Informed RRT* to belief space using the $2$-Wasserstein ($W_2$) metric. Assuming isotropic Gaussian beliefs, sampled belief states can be connected efficiently while accounting for available information and probabilistic collision constraints. This enables steering and rewiring without repeatedly propagating observations, and allows previously computed measurement information to be reused. We present a framework to generate semantically labelled digital twins for planning in real-world environments with point-cloud-based localization. Experiments in simulated environments and digital twins show faster initial solution discovery in most maps with competitive cost convergence.

#13

h=24

In CEM, a World Model Is Also a Proposal Mechanism

📅 2026-10-01cs.LGcs.RO👤 Frieder Stolzenburg (h=24)

Oliver Obst, Frieder Stolzenburg

Core Contributions

Separates the two jobs a world model does inside CEM: ranking current candidates and shaping the next iteration's sampling distribution — a scoring error affects both.
Methodologically careful: four model types generate CEM traces, every model rescores every saved candidate pool, and real-environment execution provides a ground-truth elite set.
Finds pairwise ranking agreement near chance on Walker and declining on Cheetah, even as proposal distributions contract — models steer the search more than they accurately rank it.
A pre-registered one-update intervention (swapping in environment-ranked elites once) lowers final cost in both selection and held-out units, suggesting cheap fixes targeted at proposal updates.

Show abstract

The cross-entropy method (CEM) uses world-model scores to select action sequences and fit the distribution sampled in its next iteration. A scoring error can therefore change both the present decision and the candidates considered later. We evaluate these two roles separately. Four types of predictive model generate CEM traces, and every model rescores every saved candidate pool. Executing the same candidates in the environment provides a reference elite set and proposal update. Across twelve independently trained task-seed units on Walker and Cheetah, the pre-specified proposal distance falls from the first to the final CEM iteration in every unit. Proposal widths contract and fitted means separate relative to the remaining search width. Pairwise ranking agreement stays near chance on Walker and declines on Cheetah; elite-set agreement does not improve. This comparison shows greater variation between scorers than between pool sources on Cheetah; Walker has variation in both and in their pairings. We use the original six units to select Random nonlinear for a one-update intervention, without inspecting intervention outcomes. Replacing its first model-ranked update with an environment-ranked update lowers final realised selected-sequence cost in those six units and in six further units held out from the selection.

#17

h=22

Cross-entropy optimization with prioritized constraints

📅 2026-10-01cs.RO👤 David Fridovich-Keil (h=22)

Francisco Roldan Sanchez, Pau de las Heras Molins, David Fridovich-Keil, Georgios Bakirtzis

Core Contributions

Shows that weighted-penalty optimisation can increase its score while violating the intended priority among conflicting constraints.
TierCEM filters CEM samples constraint-by-constraint from highest to lowest priority; when a constraint empties the set, it backs off and picks elites with least violation of that blocking constraint.
Removes per-constraint weight tuning while recursively guaranteeing higher-priority constraints are kept.
Tested on 2D navigation and contact-rich pushing with proprioceptive and learned world models: reversing the priority order predictably changes which constraint is relaxed, giving designers direct control.

Show abstract

When constraints conflict, an optimizer must determine which requirements to preserve and which to relax. On the one hand, a priority ordering specifies which requirements take precedence. On the other hand, penalty-based formulations encode their relative importance through numerical weights. Depending on these weights, a solution can improve its weighted score while violating intended priorities. We introduce TierCEM, a variant of the cross-entropy method that incorporates strict constraint priorities directly into elite selection without requiring per-constraint importance weights. TierCEM works by sequentially filtering sampled candidates, from highest- to lowest-priority constraint. If and when a constraint eliminates all remaining candidates, TierCEM returns to the last nonempty set and selects elites with the smallest violations of that blocking constraint, recursively preserving satisfaction of all higher-priority constraints. We evaluate TierCEM on 2D navigation and contact-rich pushing tasks in proprioceptive and learned world-model settings. Experiments show that reversing the constraint ordering changes which constraints are violated under conflict. Prioritizing progress toward the task objective also enables TierCEM to relax lower-priority constraints when they would otherwise prevent further progress.

#22

h=20

Finite-Data Safety Informativity Under Dynamic Asymmetric Actuation

📅 2026-10-01eess.SYcs.ROmath.DS👤 Abhinav Sinha (h=20)

Abhinav Sinha, Praveen Kumar Ranjan, Yongcan Cao

Core Contributions

With limited, noisy data many models remain consistent; this paper certifies whether a command is safe for all of them under asymmetric actuator limits.
Derives a support formula for the worst-case safety contribution, letting even rank-deficient data records contribute to certification.
Feasibility reduces to an affine inequality, the closest certified command to scalar root-finding, and a closed-form gate selects the largest certified fraction of a command segment.
Provides output-safety guarantees extended to all time via domain retention, demonstrated in a safety-critical vehicle study.

Show abstract

When the system model is not fully known, measurement error and limited excitation can leave several models consistent with the same finite data. A command judged safe for one model may fail for another, while limited control authority can prevent the corrective action needed to preserve safety. To ensure safety under model uncertainty and asymmetric input limits, we develop a finite-data certificate that determines whether a command can enforce a prescribed safety inequality. For a linearly parameterized safety channel with exactly known regressors and bounded aggregate residual error, we derive a support formula for the worst-case safety contribution of all data-consistent models. The formula identifies the regressor directions that admit a finite bound, allowing rank-deficient records to contribute to safety certification. Using certified componentwise bounds on actuator tracking error yields an affine inequality with a necessary and sufficient test for pointwise command feasibility. The affine inequality reduces computation of the closest certified command to a scalar root-finding problem. It also yields a closed-form gate that selects the largest certified fraction of a prescribed command segment. The proposed certificate guarantees output safety within its operating domain, provided the feedback is locally Lipschitz and the uncertainty bounds remain valid. Domain retention and full-state continuation extend this guarantee to all time. A vehicle study demonstrates that output safety can be certified from finite measurements in a safety-critical setting with model and actuator uncertainty.

Multi-Robot & Multi-Agent Coordination

h=36

Managing Context and Communication in Distributed Agentic UAV Swarms

📅 2026-10-01cs.MAcs.AIcs.LGcs.NIcs.RO👤 Marco Di Felice (h=36)

Andrea Iannoli, Ivan Zyrianoff, Angelo Trotta, Lorenzo Gigli, Marco Di Felice

Core Contributions

Studies fully distributed swarms where every UAV runs its own small language model — and identifies context management, not reasoning, as the core problem.
Represents knowledge as structured atomic notes in core/local/peer-specific memories, and disseminates them with a deterministic gossip engine that ranks by recipient-specific semantic novelty and recency.
Striking negative result: letting the SLM itself decide what to forward caused mission failure in every run, and naive flooding completed only 70–85% of runs; the proposed design completed 100%.
Roughly halves inference-token consumption vs. flooding while lowering survivor-count error in a 10-UAV search-and-rescue simulation.

Show abstract

Unmanned aerial vehicle (UAV) swarms increasingly rely on language-model agents to provide adaptive mission-level reasoning in uncertain environments. Fully distributed control, in which each UAV hosts an independent Small Language Model (SLM), removes reliance on a centralized coordinator but introduces an information-management problem: long-running interaction histories can degrade the reasoning context, while indiscriminate information dissemination increases communication and inference overhead. We address these challenges with a distributed UAV-agent architecture that enables continuous local SLM control through an event-driven reason-act-observe lifecycle. Runtime knowledge is represented as structured atomic notes and organized into core, local, and peer-specific memory. A deterministic interest-aware gossip engine selectively disseminates these notes according to recipient-specific semantic novelty and recency. We evaluate the architecture using ten UAVs in a simulated search-and-rescue mission. Our approach completes all experimental runs, whereas unrestricted flooding messages completes only 70-85\%, and delegating forwarding decisions to the SLM prevents mission completion in every run. Compared with unrestricted flooding, our approach approximately halves inference-token consumption, reduces transmitted data, and achieves lower survivor-count error.

#14

h=23

Watch, Infer, Coordinate: Inferring Robot Partner Constraints for Zero-Shot Coordination

📅 2026-10-01cs.ROcs.AIcs.MA👤 Homanga Bharadhwaj (h=23)

Suyu Ye, Zheyuan Zhang, Vaishnav Tadiparthi, Hossein Nourkhiz Mahjoub, Ehsan Moradi Pari

Core Contributions

Asks whether a helper robot can infer a partner's hidden physical limits (e.g., actuator faults) just by watching it work with another robot, then use that knowledge on a new task.
Key insight: in physically coupled tasks the healthy partner compensates, so the constraint is visible in the joint behaviour of both robots, not the constrained robot alone.
Contributes a benchmark spanning three coupled manipulation settings and an inference method that scores candidate constraints against observed joint behaviour.
Substantially improves both constraint inference and zero-shot coordination, approaching an oracle given the true constraints.

Show abstract

Robots operating in the physical world will increasingly need to coordinate with other robots, particularly in manipulation tasks where an object may be too large or heavy for a single robot to carry alone. Physical limitations caused by hardware degradation or actuator faults can restrict the actions a robot can reliably execute, yet these limitations may be unknown to its partner. We study whether a helper can infer a robot partner's physical constraints from observing it coordinate with another robot, then use the inferred capability to coordinate with the same partner on a new task. This is difficult because a demonstration shows what the constrained robot did, but not what it could have done. In physically coupled tasks, the other robot may also compensate for its limitations, making those limitations difficult to identify from the constrained robot's behavior alone. Our key insight is that these constraints shape the joint behavior of the team, making the actions of both robots informative about the constrained partner's capability. We introduce Watch, Infer, Coordinate, a benchmark spanning three physically coupled manipulation settings, together with an inference approach that scores candidate constraints using observed joint behavior. Across all three settings, our method substantially improves constraint inference and zero-shot coordination, approaching an oracle with access to the true constraints.

#21

h=21

MASkillBlender: Decentralized Whole-Body Coordination for Multi-Humanoid Loco-Manipulation via Skill Blending

📅 2026-10-01cs.ROcs.LGcs.MA👤 Guang-Hui Wen (h=21)

Yifan Hu, Luhang Hong, Mingkang Long, Danning Wang, Chengfeng Jia

Core Contributions

Coordinates multiple humanoids by learning one shared, decentralised high-level policy that blends reusable pre-trained single-humanoid skills.
Needs only task-level rewards — no task-specific motion references or heavy reward shaping, a common blocker in multi-humanoid RL.
Permutation-based data augmentation exploits robot homogeneity, with a proof that permuted samples preserve the policy-gradient direction.
Works across multiple coordination tasks and two humanoid embodiments in simulation.

Show abstract

Coordinated multi-humanoid loco-manipulation is promising yet challenging due to high-dimensional whole-body control, decentralized decision making, and scalability. While recent reinforcement learning methods have improved single-humanoid whole-body control, extending them to the multi-humanoid setting remains nontrivial and often requires substantial reward engineering or task-specific design. We propose MASkillBlender, a general multi-agent reinforcement learning framework to achieve decentralized multi-humanoid whole-body coordination. By learning a shared decentralized high-level policy over reusable pre-trained single-humanoid skills, MASkillBlender enables coordinated behaviors using only task-level rewards, without requiring task-specific motion references. To improve learning efficiency, we further introduce a permutation-based data augmentation strategy for homogeneous multi-humanoid systems, and theoretically show that the permuted samples preserve the policy-gradient direction of the original samples under the homogeneous Markov game formulation. We evaluate MASkillBlender on multiple multi-humanoid coordination tasks across two humanoid embodiments. Simulation results demonstrate that the proposed framework consistently achieves strong task performance and enables coordinated behaviors across different tasks and humanoid embodiments.

#26

h=17

OpenSpace Lab Solution to the IROS 2026 Indoor Exploration Competition

📅 2026-10-01cs.RO👤 Si-Yu Teng (h=17)

Yuxuan Zhang, Dong Li, Zezhou Sun, Yuxuan Xu, Siyu Teng

Core Contributions

Winning entry (1st, Single-Robot Public Track; 3rd in both private tracks) of the IROS 2026 indoor exploration competition.
Uses pre-trained map-completion predictions to prioritise likely unexplored areas in global planning.
A remaining-time strategy folds the return-home constraint into decisions, trading coverage against operation time explicitly.
Multi-robot mode uses shared map and intent to avoid redundant search; coverage 61.04% (public) and ~39.5–39.9% (private tracks).

Show abstract

This report presents the \textbf{OpenSpace Lab}'s solution to the Competition on Intelligent Information Gathering for Single and Multi-Robot Systems Workshops, organized as part of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) 2026. Our team reached 1st place in the Single-Robot Public Track and 3rd place in both the Single- and Multi-Robot Private Tracks. The single-robot framework utilizes pre-trained map completion predictions for global planning to prioritize unexplored areas. To reconcile map coverage with limited operation time, we introduce a remaining-time-based exploration strategy that integrates homing constraints into the decision-making process. For multi-robot exploration, we utilize a utility-driven target selection strategy that balances observation gains, movement costs, and budget constraints, leveraging shared map and intent data to eliminate redundant search and maximize coordination efficiency. Our solution reached a 61.04\% coverage rate in the Single-Robot Public Track, while reaching 39.53\% and 39.91\% coverage in the Single- and Multi-Robot Private Tracks, respectively. An extended full-length paper based on this report is currently being prepared for submission, and the source code will be released upon acceptance of the full manuscript at https://github.com/OpenSpace-Lab/Indoor-Exploration-IROS2026.

Simulation, Digital Twins & 3D Assets

#15

h=23

H-SPAR: Hydrodynamic-aware Simulation for Particle Transport and Autonomous Robots

📅 2026-10-01cs.RO👤 Sajad Saeedi (h=23)

Navid Zarrabi, Nariman Yousefi, Sajad Saeedi

Core Contributions

Couples time-varying flow fields, Lagrangian particle transport, probabilistic sampling and ROS 2/Gazebo USV autonomy in one simulator — prior marine simulators model these separately.
The same precomputed flow drives both particle advection and current-induced forces, so mission cost and sampling yield are evaluated under consistent physics.
Reveals a planning-execution gap: current-aware SVF-RRT* saves 69.4% upstream cost at planning time but only 41.7% when executed in time-varying currents.
Sweep orientation alone changes particle-sampling rate by up to 22.2% — showing that coverage plans must be designed against the flow, not just the map.

Show abstract

Environmental robotic sampling requires considering the dual influence of water currents on robotic motion and particle transport. Existing marine robotics simulators generally model flow, autonomy, and sampling targets separately, limiting joint evaluation of mission cost and sampling performance. H-SPAR integrates spatially and temporally varying velocity fields, Lagrangian particle transport, probabilistic sampling, and ROS 2/Gazebo-based uncrewed surface vehicle (USV) autonomy. In this work, shared precomputed flow fields drive particle advection and current-induced forces during closed-loop vehicle execution. Path-planning experiments show that the existing current-aware planner SVF-RRT* achieves 69.4% lower upstream cost than conventional RRT* at the planning level, but this reduction falls to 41.7% during execution under time-varying currents, reflecting temporal flow variation, vehicle motion constraints, and path deviation omitted during planning. Coverage experiments show that sweep orientation changes the particle-sampling rate by up to 22.2% under the complete H-SPAR configuration. These findings highlight the importance of evaluating planning, vehicle execution, particle transport, and sampling together under consistent hydrodynamic conditions. The project webpage is available at https://sites.google.com/view/h-spar, and the open-source code is available on GitHub at https://github.com/naviiidz/h-spar-sim.

#28

h=16

3DROID: A Renderable 3D Gaussian Dataset with Measured Per-Scene Reliability

📅 2026-10-01cs.CVcs.RO👤 N. Kwak (h=16)

Wonguen Cho, Junhoo Lee, Nojun Kwak

Core Contributions

Shows that feed-forward 3D Gaussian splatting optimised for photometric fidelity can lose metric scale when camera extrinsics are unreliable — a problem for using 3DGS as manipulation supervision.
Finds that pose conditioning helps novel-view fidelity, but its geometric benefit depends on how reliable the injected extrinsics are.
Proposes a calibration-aware pipeline that anchors reconstructions to the robot's metric workspace.
Releases 3DROID, a renderable, metric-anchored dataset with per-scene reliability scores so users can filter data by trustworthiness.

Show abstract

Robot manipulation models primarily reason from 2D observations while acting in the 3D physical world. To bridge this gap, recent work has augmented robot data with geometric priors such as depth, point clouds, and 3D trajectories, while renderable 3D Gaussian representations provide another promising form of 3D supervision. However, 3DGS representation is designed mainly for photometric fidelity and may not preserve real-world metric scale, particularly when the supplied camera extrinsics are unreliable. We study the effect of extrinsic reliability and pose conditioning on feed-forward 3DGS, and propose a calibration-aware pipeline that anchors reconstructed scenes to the robot's metric workspace. Our experiments show that pose conditioning improves novel-view fidelity, while its geometric benefit depends on the reliability of the injected extrinsics. Using this pipeline, we present a renderable, metric-pose-anchored dataset with scene-level reliability information for robot manipulation research. Our dataset is available at https://huggingface.co/datasets/wonguen/3DROID

#29

h=14

TouchTherm: Building Multimodal Digital Twins of Objects for Tactile and Thermal Rendering

📅 2026-10-01cs.RO👤 top author unresolved (h=14)

Yitao Zhang, Hong Ying, Haoran Guo, Xiaoying Zhou, Guanyu Chen

Core Contributions

Builds object assets with three senses — visual, tactile and thermal — whereas existing 3D datasets capture geometry and appearance only.
Photometric-stereo normals registered to structured-light scans yield micro-height fields for optical tactile rendering, while a coarse mesh handles collisions.
Multiview infrared video of natural cooling reconstructs a physics-regularised dynamic thermal field: held-out MAE 0.465 °C at 30 s and 0.592 °C at 45 s across 20 objects.
Assets enable synthetic-to-real tactile object recognition and drive a glove-based VR display with spatially and temporally varying heat.

Show abstract

Robotic simulation and virtual reality increasingly require object assets that capture not only visual geometry but also the physical cues underlying tactile and thermal interaction. Existing 3D datasets and reconstruction methods primarily represent object-scale geometry and visual appearance, overlooking microscale surface structure for high-fidelity haptic rendering and transient temperature dynamics for temperature-aware interaction. We present TouchTherm, a framework for constructing simulation-ready visuo-tactile-thermal object assets from real-world objects. For visual and tactile reconstruction, we combine structured-light scanning with multiview normal maps obtained from photometric stereo. The normal maps are registered to the scanned geometry and transformed into tangent space to recover local micro-height fields for optical tactile rendering, while the coarse mesh handles collision detection. For thermal reconstruction, we capture synchronized multiview infrared videos of natural cooling following controlled heating and reconstruct a physics-regularized dynamic thermal field. Experiments on 20 objects show that the reconstructed micro-height fields preserve dominant surface structures and recover higher-frequency details beyond the coarse geometry, while the thermal fields achieve held-out surface-temperature MAEs of 0.465 degrees C and 0.592 degrees C at 30 s and 45 s, respectively. The resulting tactile assets support synthetic-to-real object recognition from tactile observations, while a glove-based VR system demonstrates spatially and temporally varying thermal feedback. These results highlight the potential of TouchTherm for multimodal sensory simulation and temperature-aware virtual interaction.

Field, Medical & Human-Assistive Robotics

h=29

Towards Physical Underwater Robotic Assistance for Scuba Diver Movement in Confined Spaces

📅 2026-10-01cs.RO👤 Junaed Sattar (h=29)

Demetrious T. Kutzke, Junaed Sattar

Core Contributions

A wearable robot that gives divers thruster-generated lateral directional guidance — distinct from prior exoskeletons that provide propulsion — to keep safe distance from reefs, rigs or ordnance.
Fuses monocular depth estimation with force feedback through submersible thrusters, turning perception into a haptic cue the diver feels rather than a display they must read.
IRB-approved in-water studies with eight divers in pool and ocean show only ~10% of max thrust is enough to steer a human's movement.
Opens a new category of physical human-robot interaction underwater, where visual/acoustic communication is limited.

Show abstract

Scuba divers are taught to control their depth to avoid rapid ascents and descents, which could result in serious injuries such as gas embolisms and barotrauma. However, many underwater tasks necessitate lateral control, maintaining distance between subsea structures such as coral reefs, submerged drilling instrumentation, or unexploded ordnance. In this work, we discuss a first-of-its-kind wearable robotic solution providing thruster-actuated directional guidance to a diver, as distinct from prior propulsive-assistance exoskeletons. We introduce ``Robotic Assisted Diver Movement in Confined Spaces'' (RADMCS), a wearable robot that assists divers in maintaining a fixed distance from subsea structures by leveraging perception techniques in monocular depth estimation and force-feedback from submersible thrusters to provide haptic feedback. Its small and compact form factor creates a foundational platform that could be expanded to include more sophisticated control and navigation behaviors. We present results from Institutional Review Board (IRB) in-water studies with eight human scuba diver participants on threshold sensitivity tests in both a closed-water swimming facility and ocean environments; distance-maintaining experiments in a closed-water facility; and form, fit, and function testing in the ocean. We demonstrate that relatively low thrust values (10 percent of maximum) allow robotic direction of a human's movement using the physical sensation of the robot's guidance.

h=29

Ex vivo breach detection using electrical conductivity during robotic pedicle drilling in the spine

📅 2026-10-01cs.RO👤 Q. Grimal (h=29)

Jorge Andrés Pérez Velásquez, François Teyssere, Thibault Chandanson, Quentin Grimal, Brahim Tamadazte

Core Contributions

Replaces intraoperative CT and registration with a sensor already inside the drill: a proprioceptive electrical bioimpedance probe (SpineGuard) on a robotic pedicle-drilling tool.
A real-time algorithm watches for abrupt conductivity changes that precede a breach toward the spinal canal and stops the drill before it happens.
Prevented breaches in all 51 ex-vivo drilling cases, with no external sensors or imaging.
Significant because it reduces radiation exposure and registration complexity in scoliosis surgery — safety comes from tool-tissue physics rather than image alignment.

Show abstract

Purpose: Pedicle screw placement is technically demanding in scoliosis treatment. High precision is required due to limited visibility, anatomical variability, and the risk of complications. Although robotic systems assist CT-based planning and execution, they still rely on ionizing intraoperative imaging and complex registration. This study proposes robotic pedicle drilling with real-time preventive breach detection using electrical bioimpedance sensing. Methods: We developed a robotic approach combined with a pedicle-drilling tool equipped with a proprioceptive electrical bioimpedance sensor developed by SpineGuard. A real-time detection algorithm was designed to analyze the electrical bioimpedance signal during drilling and identify abrupt changes in conductivity associated with potential breaches towards the spinal canal. The method operates without external devices or sensors. Results: The ex vivo experiments showed that the proposed method prevented breaches in 100 of the 51 drilling cases. These findings demonstrate the system's ability to detect potentially hazardous events during drilling and to stop the procedure before. The ex vivo experiments demonstrated that the proposed method prevented breaches in all 51 drilling cases. Conclusions: This work demonstrates the feasibility of robotic pedicle drilling with electrical bioimpedance sensing for real-time breach prevention. Using only the tool signal, the method eliminates the need for external sensing systems and supports safer pedicle screw placement.

#10

h=26

ALFRED: Requirement-driven development of an open-source mobile manipulator for long-term plant monitoring

📅 2026-10-01cs.ROcs.ARcs.CV👤 F. A. Cheeín (h=26)

Ciarán Miceal Johnson, Christopher Quail, Garry Ellard, Alistair McConnell, Steve Tonneau

Core Contributions

An open-source six-DoF mobile manipulator for repeated outdoor plant monitoring, released with design files — addressing the fact that robots behind long-term field datasets are rarely reproducible.
Documents four builds driven by six explicit requirements (durability, modularity, repairability, sensing reach, endurance, reproducibility), quantifying how the robot's own frame limits arm reach and occludes sensors.
Usable arm-reachable poses rose from 34.0% → 60.0% → 66.1% across builds; ray casting shows only the final build keeps the LiDAR clear front and back.
Completed a year of monthly forest surveys (528 traversals) without a missed collection, with ~6 h build swaps — and highlights that the robot description file must evolve with the hardware.

Show abstract

Tracking seasonal change in crops and forests requires observing the same plants repeatedly. Ground robots can do this at close range, and a manipulator gives their sensors more viewpoints. Yet the robots behind long-term field datasets are rarely released with their design files, and how a robot's own structure limits arm reach and occludes its sensors is seldom compared between builds. We present ALFRED, an open-source mobile manipulator built from commercially available components. It carries a six-degree-of-freedom arm, LiDAR, RGB-D cameras, RTK GNSS and an IMU on an Ackermann-steered base, all mounted on a reconfigurable aluminium strut frame, and runs containerised ROS software. It was developed through four builds against six requirements for repeated outdoor deployment: durability, modularity, repairability, sensing reach, endurance and reproducibility. Model-based analysis of the last three builds shows the usable share of the arm's reachable poses rising from 34.0% to 60.0% and then 66.1%, and ray casting shows that only the final build keeps the frame-mounted LiDAR's horizontal view clear both forwards and backwards. ALFRED completed a year of monthly forest surveys (528 traversals) without missing a scheduled collection. This was despite battery degradation, reconfiguration for another researcher's study, and the parallel development of ALFRED 2.0 for autonomous crop-row operation, with each switch between builds taking about six hours. The deployment also showed that mechanical modularity is only as dependable as the robot description that tracks it.

#18

h=22

ColoACT: Multi-Cue Action Chunking for Smooth Autonomous Colon Navigation on a Self-Propelled Endoscopic Robot

📅 2026-10-01cs.RO👤 Chao-Yang Shi (h=22)

Jian Hu, Shujing He, Leixin Chang, Zongze Li, Ding Huang

Core Contributions

An Action Chunking Transformer policy for a compact self-propelled bevel-gear endoscopic robot, aimed at reducing loop formation and tissue trauma in colonoscopy.
Augments RGB with estimated relative depth and a gradient-based pseudo-elevation map that makes fold ridges salient — targeting the weak texture and specularity that break pure-vision depth.
Overlapping action chunks fused by temporal ensembling give smooth continuous motion instead of the switching logic of geometric pipelines.
In ex-vivo porcine colons: 85.4% (straight) and 72.5% (curved) success, 70% on 90° turns, 60% on double bends, with triple-bend feasibility shown.

Show abstract

Autonomous colonoscopic navigation can reduce operator burden and the risk of loop formation or tissue trauma, but remains challenging due to deformable anatomy, weak-texture and specular endoscopic visuals, and contact-rich viscoelastic interactions. Existing methods either rely on geometry-driven pipelines, which are efficient and interpretable yet brittle due to manually engineered features and switching logic, or adopt learning-based policies, whose inferred depth/geometry can become temporally inconsistent or overly smooth under weak texture and specular highlights while simulation-trained variants (e.g., deep reinforcement learning) may further suffer from a sim-to-real gap. We propose ColoACT, an autonomous navigation system that integrates an RGB-D-E based Action Chunking Transformer policy (ColoACT policy) for a compact self-propelled Bevel-Gear-Based Endoscopic Robot (BGER). The ColoACT policy augments RGB with estimated relative depth and a gradient-based pseudo-elevation map to enhance fold-ridge saliency and other high-frequency geometric cues, and enables smooth continuous control of the BGER by predicting overlapping action chunks and fusing them via temporal ensembling. In different \textit{ex-vivo} porcine colons (approximately 60 cm), our system achieves success rates of 85.4\% and 72.5\% in straight and curved segments, respectively, and achieves 70\% success in 90-degree turns and 60\% in double-bend sequences, with feasibility further demonstrated in challenging triple-bend segments. The project page is available at: https://Adamhu1.github.io/ColoACT/.