Today's batch of 30 papers clusters around three dominant currents. The largest is a continued push toward data-efficient manipulation learning: PhysClaw-0 cuts human teleoperation time to 16% via a reusable Corrective Memory, Reverse to Advance exploits the fact that undoing a hard task is often trivially easy to manufacture cheap supervision, and the Industrial Dexterity Benchmark shows a multimodal diffusion policy reaching 78% success from only ~100 demonstrations per task phase. All three attack the same bottleneck — teleoperation cost — from different angles (memory reuse, structural task asymmetry, and multimodal sensing efficiency), suggesting the field is converging on the view that smarter data curation, not just more data, is what closes the manipulation generalization gap.
A second current concerns how much of a robot foundation model actually needs to survive finetuning or run at inference. GigaWorld-Policy-0.5 shows that the future-video-generation step in World Action Models can be dropped at inference (85ms on a single RTX 4090) without losing its training-time benefit; Kepler-Encoder-v0.1 shows that fusing force and proprioception into training makes a vision-only latent recover state information the pixels alone don't carry; and Anchor-Align shows that behavior-cloning finetuning silently erodes pretrained VLM representations unless explicitly anchored against a frozen copy. Read together, these papers argue that the expensive parts of multimodal pretraining (video prediction, force sensing, language grounding) are separable from what needs to run at test time — you can train with everything and deploy with less, provided you explicitly protect what would otherwise be thrown away.
A third, quieter theme is safety folded into learning rather than bolted on afterward: Discriminative Barrier Functions constrains IRL's reward search to the space of Control Barrier Functions so the recovered policy inherits a formal safety certificate, while Safe Overtaking uses RL to tune only the CBF decay parameter online rather than learning a full control policy. Both keep the safety-critical structure classical and let learning handle just the part that needs adapting. That restraint shows up again in the batch's emphasis on physical validation: APT-RL's quadruped hits 6 m/s drop-downs outdoors, WNOJ-LIO is tested on a racing car up to 208 km/h, EgoHTR trains a real Unitree G1, and the DVOP orienteering solver and Layered Risk Mapping wheelchair framework both close with real hardware deployments rather than simulation alone — a batch that is notably unwilling to stop at simulated results.
LiDAR SLAM map consistency, aerial GNSS-denied localization, fiducial markers, spiking-network place recognition, and hybrid model/learning tracking filters.
Data-efficient approaches to dexterous and deformable manipulation that attack the teleoperation-cost bottleneck via corrective memory, task-reversal, hybrid physics-learning simulators, and multimodal sensing.
Intersection decision-making, racing overtakes, heavy-truck planning, and hybrid vehicle-in-the-loop validation for automated and racing vehicles.
Efficient action-centered world models, multimodal state encoders, and representation-preserving VLA finetuning β all racing to capture pretraining and future-prediction benefits without paying their usual inference or generalization costs.
Language-driven goal navigation, UAV orienteering under real dynamics, multi-agent 3D reconstruction, and RL-based navigation through unsteady flows.
Terrain-aware data capture and multi-skill locomotion policies pushing quadrupeds and humanoids toward robust, high-speed real-world traversal.
Safety certification embedded directly in learning (barrier functions, adaptive CBFs), risk-aware planning, and non-verbal communication of robot uncertainty.
High-quality maps are fundamental for robotics tasks such as navigation and planning. Although modern graph-based LiDAR SLAM systems achieve good trajectory accuracies, a low trajectory error alone does not guarantee geometrically consistent maps, particularly at revisit locations where missed loop closures and residual drift can produce local misalignments. In this work, we address the problem of jointly improving global trajectory estimation and local map quality in 3D LiDAR SLAM. We first propose a framework to efficiently estimate geometry-dependent information matrices for ICP, enabling principled weighting of odometry constraints in a pose graph. We then introduce a hierarchical loop-closure module that decouples place recognition from geometric registration, together with a retroactive loop-closure module that exploits the optimized pose graph to recover missed loop closures. We also propose an evaluation protocol to measure map consistency at revisit locations. We evaluate our SLAM system on several datasets against state-of-the-art LiDAR SLAM systems. Experimental results demonstrate global trajectory accuracies on par with or better than existing methods while consistently improving local geometric map consistency at revisit locations.
Unmanned Aerial Vehicles increasingly rely on visual fiducial markers for autonomous navigation and precision landing. Standard markers suffer from limited operational ranges, and existing recursive/fractal markers either require the marker's center to remain visible or are limited in recursion depth. We propose a novel Recursive ArUco marker design that transforms any standard fiducial marker into a recursive marker with arbitrary depth by embedding complete markers within both the black and white bits of the parent marker via a modified bit-sampling strategy, guaranteeing unlimited recursion depth and robust detection even with partial occlusion. Maintaining a single unique identifier across all recursive scales provides an extensive dictionary of multiple unique landing pads for simultaneous multi-UAV operation.
We present AeroMap3D, a monocular 6-DoF UAV localization system that anchors onboard imagery to visual, geometric, and semantic map priors for GNSS-denied navigation. AeroMap3D addresses the cross-view discrepancy between UAV imagery and satellite maps and the structural inconsistency between bare-earth DEMs and urban scenes via a lightweight adapter enabling reliable UAV-to-map registration without finetuning, and by lifting 2D UAV-map correspondences onto DEM terrain while using OpenStreetMap annotations to reject unreliable matches before RANSAC-PnP pose estimation. Delayed map-based pose measurements are fused with relative-motion priors using a delayed-state EKF. Without retraining, AeroMap3D localizes all trajectories across eight Austin sites within 50 m and achieves 5.88 m mean 3D error over 55 km of flight.
Spiking Neural Networks trained through unsupervised STDP have been explored for visual loop closure, but state-of-the-art models fail to reach the high Recall at 100% Precision needed for reliable autonomous navigation. We present a discrete, tensor-native implementation of the STDP-based SNN-VPR pipeline using PyTorch with snnTorch, evaluated on a 100-place Nordland dataset. We show neuron assignment with a closed-form, deterministic tensor pipeline provides significantly higher R@100P than standard argmax. State reset after each query helps regardless of assignment method, and velocity-compensated sliding window aggregation over 5 frames reaches R@100P = 100.00% for constant-velocity traversal at an additional 0.20 ms latency.
Maneuvering target tracking in three-dimensional space remains challenging due to complex motion dynamics and model mismatch. We propose IMMNet, a hybrid model/data-driven algorithm integrating the interpretable structure of the interacting multiple model (IMM) algorithm with learnable neural components. Unlike end-to-end black-box methods, IMMNet preserves the Bayesian inference mechanism essential for real-time radar applications while adaptively learning motion patterns and noise characteristics from data. Extensive experiments demonstrate IMMNet consistently outperforms existing algorithms across various scenarios.
LiDAR-inertial odometry (LIO) is key to autonomous navigation, but high-dynamic driving exposes intra-scan motion distortion and vibration-contaminated inertial measurements. Most real-time pipelines propagate state by integrating raw IMU measurements, propagating inertial noise into both the corrected scan and scan-to-map registration. We present WNOJ-LIO, a LiDAR-IMU fusion framework based on a White-Noise-on-Jerk EKF, employing a decoupled WNOJ prior for state prediction and treating the IMU as a high-frequency measurement source rather than the driver of state propagation. Real-world experiments on an autonomous racing car across four segments with speeds from 53 to 208 km/h validate the method. Source code is publicly available.
Surgical tissue retraction requires effective manipulation planning under partial and noisy perception. We study state estimation for deformable tissue retraction, where only sparse observations of the tissue surface are available at decision time. We propose a learned state estimator that reconstructs the full deformable mesh state from 40 noisy vertex observations. The estimator combines a multilayer perceptron with a low-dimensional PCA latent representation and is trained using geometry-aware regularization that encourages smooth and physically plausible deformations. We evaluate the approach in a 2D deformable sheet simulation using single-step and multi-step retraction planning. Results show that the learned estimator achieves 98.1% of oracle performance in multi-step retraction while supporting efficient inference.
Real-world deployment of embodied agents requires active exploration, visual grounding, and interactive intent disambiguation. We present REAL, an agentic framework for open-world mobile manipulation that establishes sim-to-real-consistent environment APIs without oracle perception and integrates a simulated user for human-in-the-loop interaction. We introduce REAL-Bench, a benchmark spanning 241 tasks. Our trained agent outperforms leading commercial closed-source VLMs on interactive tasks with a 56.9% success rate, and achieves a 78.3% end-to-end success rate over 60 real-world episodes on a physical dual-arm mobile robot, demonstrating robust zero-shot transferability to unseen household scenarios.
Autonomous data collection governs the volume and quality of real-world trajectories for manipulation policy learning. Existing pipelines reduce human effort via self-resetting, VLM verification, or language-guided correction, yet episode-scoped fixes must be reissued whenever the same failure recurs. We present PhysClaw-0, a human-robot symbiotic agentic system in which corrections are retained and reused across rounds via an LLM parser mapping each utterance to a structured adjustment stored in Corrective Memory. On a real-robot desktop-clearing testbed, PhysClaw-0 matches teleoperation episode success while reducing human working time to 16%, and raises average single-attempt success from 12.5% to 47.5%.
Dexterous manipulation remains a critical bottleneck in industrial automation. This work presents a progression from classical, modular robotics pipelines toward an end-to-end multimodal imitation-learning framework for industrial dexterous manipulation, introducing Industrial Dexterity Benchmark (IDB) boards mimicking datacenter cable management, automotive cable harnesses, and gearbox assembly; a scalable imitation learning framework (DAG-ROS); and a multimodal diffusion-based policy (AG-iDP3) fusing RGB, point clouds, joint positions, and wrist-frame wrench data. The best configuration, a multimodal Diffusion Policy with an R3M encoder, reaches a 78% grasp-and-insert success rate, versus 36% for a single-camera RGB baseline, using only ~100 teleoperated demonstrations per task phase.
High-quality teleoperation datasets are costly to collect, particularly for hard tasks. Many tasks exhibit directional asymmetry: completing the forward hard task is difficult, whereas reversing it is comparatively easy. We propose a teleoperation-cost effective framework for hard policy learning via temporal reversal of easy tasks, consisting of a closed-loop data collection pipeline alternating hard-task and easy-task policies, a hierarchical data refinement pipeline temporally inverting easy-task rollouts and filtering low-quality motion, and an iterative policy learning method. Across two simulated benchmarks and real-robot experiments, our method improves hard-task success rates with higher data efficiency and more stable training compared to reversal-based and RL baselines.
Simulating deformable objects is essential for robotic manipulation applications, yet accurately predicting their dynamics remains challenging. We propose Physics-Guided Residual Dynamics (PGRD), a hybrid simulation framework combining an optimizable spring-mass simulator as a backbone with a learned neural network predicting residual corrections. We adopt a velocity-based formulation for stable simulation and a sliding-window transformer to capture temporal dependencies. PGRD produces more accurate results than both purely physics-based and learning-based methods on diverse real-world deformable objects, and we demonstrate utility in manipulation planning via MPC and interactive simulation via action-conditioned video prediction with 3D Gaussian Splatting.
European safety regulation now permits a large share of automated-driving homologation evidence to be produced virtually, provided a validated physical-virtual facility generates it. We present a deployed hybrid Vehicle-in-the-Loop platform that couples a real instrumented vehicle with a CARLA-based digital twin through a V2X message pipeline, and report its first integrated operation on a public-road-representative test track. A real vehicle streams ETSI-compliant CAM/CPM messages into the digital twin, where a GPU-accelerated Cooperative Perception module fuses them into a probabilistic occupancy grid during scenario runtime. Results show CP substantially widens field-of-view coverage and improves occupied-cell recall, and that beyond a moderate localization-noise threshold, positioning uncertainty, not weather, becomes the dominant error source.
This paper presents a novel framework for integrated Decision-Making and Trajectory Planning for automated vehicles at unsignalized intersections. The approach leverages a Finite Horizon Optimal Control Problem that employs Time-Varying Artificial Potential Fields. By utilizing short-horizon motion prediction and a dedicated conflict-zone occupancy coefficient, the framework accounts for potential collisions within the FHOCP. Simulation results in multi-vehicle traffic scenarios demonstrate the effectiveness of the approach.
Learning-based planners' performance on heavy-duty trucks, particularly distributed electric-drive trucks (DETs), remains largely unexplored. This letter presents nuTruck, the first high-fidelity benchmark designed to support large-scale neural network training and closed-loop evaluation for DETs. We incorporate a highly accurate nonlinear truck dynamical model enabling independent driving and steering of all wheels and capturing dynamic load transfer, allowing quantitative assessment of rollover risk in closed-loop simulation. We adapt several rule-based and learning-based planners as baselines and, using nuPlan driving scenarios, conduct extensive closed-loop evaluations of both collision-free planning and dynamical safety.
Autonomous racing overtaking requires balancing competitive performance with safety under nonlinear vehicle dynamics and real-time constraints. MPC combined with Control Barrier Functions provides forward invariance of a safe set, but fixed-decay discrete-time CBFs can become overly conservative. This paper proposes a hierarchical overtaking framework separating maneuver-level decision making from safety-certified trajectory control. A high-level MIQP resolves passing-side selection, while a nonlinear Frenet-frame MPC enforces dynamics and safety through embedded CBF constraints. A reinforcement learning policy adapts the CBF decay parameter online. Simulation and scaled-hardware experiments show the adaptive strategy attains the highest aggregate success rate without per-track tuning.
Autonomous racing demands planning algorithms that balance vehicle dynamics with strategic decision-making in competitive multi-agent scenarios. Directly solving full dynamic games online is computationally prohibitive. This paper proposes a hybrid architecture integrating game-theoretic reasoning into a sampling-based motion planner. Building upon an Ξ±-potential game formulation, we utilize an offline-learned potential function to capture multi-agent interactions, and a gradient-based optimization to generate an Interaction Reference Path that serves as a dynamic cost bias within a high-frequency sampling planner. Results on the Yas Marina Circuit demonstrate the approach successfully induces defensive behaviors like blocking without the computational burden of full dynamic game solvers.
World Action Models (WAMs) improve robot policy learning by jointly modeling actions and future visual observations. A common design in existing WAMs is to explicitly generate future videos at inference time, incurring substantial computational overhead. GigaWorld-Policy addresses this with an action-centered formulation where future visual dynamics are used during training while action-only decoding is used at inference. GigaWorld-Policy-0.5 introduces a Mixture-of-Transformers architecture separating visual dynamics modeling and action generation into specialized experts, achieving 85ms inference latency on a local RTX 4090. An agent-based AutoResearch pipeline systematically searches training configurations.
A robot must understand the state of its own body, but a camera sees only part of it. We present Kepler-Encoder-v0.1, a robot-first multimodal encoder that fuses vision, proprioception, and force/torque into a single shared latent, trained self-supervised by masked cross-modal prediction under the LeJEPA/SIGReg objective. At evaluation only vision enters. On the RH20T corpus, the vision-only latent recovers end-effector state, and force in particular, significantly above raw frozen-ViT features on held-out scenes. A single embodiment-agnostic encoder covers four robots. The frozen latent's own cross-modal prediction error is a training-free invalid-state monitor (AUROC 0.90 on out-of-range states).
Driving-world generation has emerged as a core capability for scalable autonomous-driving simulation, yet existing methods remain limited in object-level controllability and long-horizon stability. We present M4World, a multi-view and multimodal generative driving world model that synthesizes future surround-view video streams and synchronized LiDAR scans while supporting interactive object manipulation and stable minute-long streaming. Stable minute-long streaming is achieved through a multi-stage training framework enabling online causal generation in only four denoising steps. We further introduce an automated VLM-based judging pipeline that evaluates scene-level condition adherence, view-wise object controllability, and cross-view object consistency.
Finetuning a pretrained VLM on robot demonstrations via behavior cloning (BC) has become standard for VLA policies, but BC finetuning progressively overwrites pretrained representations. Co-training on web image-text data does not prevent this since it applies language and action losses to separate observations. We propose Anchor-Align: Vision-Language Anchoring distills layer-wise representations from a frozen VLM copy, while Language-Action Alignment converts action targets into discrete motion-direction labels trained jointly with action prediction on the same observation. On a physical xArm7, Anchor-Align improves real-robot success on both tested architectures (28% to 54% and 37% to 60%), with consistent gains on LIBERO-PRO, LIBERO-Plus, and CALVIN.
This paper introduces the Dynamical Vehicle Orienteering Problem (DVOP), a generalization of the Orienteering Problem (OP). The OP maximizes the reward collected from spatial targets under a limited travel budget; the DVOP extends it by accounting for both external and vehicle-actuated forces. We study the DVOP in the context of multi-rotor UAV flight planning, using a three-dimensional Point-Mass Model constrained by maximum velocity and acceleration and subject to gravity, with the travel budget expressed as maximum flight time. We propose two solution approaches: a Branch-and-Bound procedure combining Non-Linear Programming and Mixed-Integer Linear Programming, and a Large Neighborhood Search metaheuristic. Experiments on benchmark instances show improvements of up to 37% over state-of-the-art solutions for the Kinematic Orienteering Problem, and a real-world deployment on a multi-rotor UAV verifies the proposed trajectories.
Natural language interaction provides an intuitive way for non-expert users to communicate with robotic platforms. This work presents a language-driven navigation framework that enables mobile robots to interpret user requests in natural language and autonomously navigate towards a destination. The framework is composed of modular ROS 2 components that transform natural language instructions into navigation actions via the Nav2 stack. The system is evaluated in simulation and real-world scenarios using a TurtleBot3 Waffle and a Unitree Go2 robot with a RealSense camera, and successfully interprets both direct commands and contextual requests.
Active 3D reconstruction requires selecting informative viewpoints under limited sensing budgets. In multi-agent settings, coordination inefficiencies such as redundant observations and spatial clustering can significantly reduce reconstruction quality. We present COLMAR, a cooperative view policy learning framework formulating viewpoint allocation as a shared policy optimization over map-centric observations with a reconstruction-aware objective. The policy is trained using parameter-sharing PPO with independent per-agent action selection at deployment, conditioned on a fused team map without inter-agent message passing. Experiments on GLEAM and Replica demonstrate up to 54% higher reconstruction accuracy and 49% greater coverage over baselines.
Autonomous robotic navigation in nonstationary time-varying fluid flows remains a fundamental challenge due to partial observability. Classical optimal control frameworks require unrealistic a-priori global flow knowledge, while biological systems navigate successfully using localized sensory cues. We present a reinforcement learning approach using TD3 to train agents to reach arbitrary targets within a parametric, chaotic double-gyre flow, evaluating five bio-inspired observation strategies. An agent that senses and remembers a set number of flow velocity measures achieves the highest performance. Velocity-aware agents optimize energy efficiency, whereas vorticity sensors provide superior structural mapping. Incorporating explicit global flow parameters decreases navigation performance.
Deploying humanoid robots in unstructured terrain remains an open problem. Existing dataset pipelines fail to capture human-scene sequences in challenging environments. We introduce the Egocentric Human-Terrain Reconstruction (EgoHTR) dataset, an open-sourced reconstruction pipeline capturing 55 scene-aligned 4D human motion sequences in diverse, complex environments using a multi-sensor setup of egocentric wearables and a portable 3D scanner. The dataset comprises over 150k frames, evaluated against motion-capture ground truth demonstrating state-of-the-art accuracy. We leverage this data to train perceptive locomotion policies, demonstrating hardware deployment on a Unitree G1.
Enabling quadrupedal robots to traverse complex terrains requires seamless integration of multiple motor skills and high-speed perceptive locomotion using only onboard sensors. We present APT-RL, a unified framework enabling multi-skill locomotion through autonomous skill transitions utilizing only onboard perception and computation. Our approach generates large-scale, feature-rich 2D motion datasets through trajectory optimization with simplified dynamics, enabling training of diverse, reusable locomotion skills that transfer to a real quadruped operating on complex uneven terrains and extend naturally to 3D environments. Real-world experiments demonstrate agile maneuvers including dynamic drop-downs reaching instantaneous peak speeds of up to 6 meters per second across stairs, hurdles, stepping stones, gaps, and fallen branches.
Inverse Reinforcement Learning (IRL) algorithms are powerful tools for learning from and generalizing expert demonstrations, but they often rely on unconstrained exploration, rendering them unsafe for real-world deployment. Control Barrier Functions (CBFs) can guarantee safety, but analytical design can be time-consuming. This work constrains reward function candidacy during IRL to the space of CBFs, enabling the data-driven recovery of barrier functions directly from unlabeled expert observations. We demonstrate the recovered barrier function is robust to unsafe states entirely absent from expert data, and benchmark against standard IRL baselines in simulated navigation and a real-world obstacle avoidance task.
Robots operating in human-robot collaboration must communicate uncertainty arising from incomplete or ambiguous perception. This work introduces a mathematical framework for expressing perceptual uncertainty through robotic manipulator motion. Drawing on Laban Movement Analysis, robot behavior is organized in a Commitment-Vigilance state space that maps uncertainty-related states to distinct Laban Effort signatures. Five motion primitives are parameterized using eleven kinematic and geometric descriptors. A video-based human-subject study evaluated recognition of four expressive trajectories and the influence of individual descriptors on perceived intensity; participants reliably identified the intended behavioral states.
In expeditionary medical facilities, routine patient transport imposes a compounding burden of PPE consumption, staff diversion, and elevated infection risk under surge conditions. We present a layered risk mapping framework that fuses four heterogeneous environmental hazards (terrain slope, static and dynamic obstacles, semantic traversability) into a unified probabilistic cost surface via a Noisy-OR fusion model. In a paired Monte-Carlo evaluation, risk-informed fusion reduces collision rates from over 73% to under 32% and more than doubles obstacle clearance relative to a risk-unaware baseline. We validate the framework on a commercial powered wheelchair across three mission profiles in indoor and outdoor deployments.