MEKAWIRE
Research

Papers

New robotics and physical-AI papers, with what each one means for the industry.

276 papers
TopicAllManipulation · 27VLA · 22Perception · 12Robotics · 11Navigation · 10Humanoid · 9Learning · 8Safety · 5World Action Models · 4Simulation · 3Aerial manipulation · 2Dexterous manipulation · 2Sim-to-Real Manipulation · 2Social Navigation · 2Vision-Language Navigation · 2World models · 2World-action model · 23D Assembly · 13DGS dataset · 15G Simulation · 1AMR Operations · 1Active vision manipulation · 1Adaptive MPC · 1Adaptive action horizons · 1Adaptive intervention · 1Aerial · 1Aerial Manipulation · 1Aerial navigation · 1Agricultural Robotics · 1Articulation Estimation · 1Autonomous Driving · 1Autonomous Exploration · 1Battery-aware RL · 1Belief-space planning · 1Bimanual Manipulation · 1Byzantine localization · 1Canal inspection · 1Co-speech Humanoid · 1Compositional Manipulation · 1Construction navigation · 1Contact-Rich Manipulation · 1Contact-rich Manipulation · 1Continuum sensing · 1Curriculum RL · 1Dexterous Manipulation · 1Distributed Manipulation · 1Distribution Shift · 1Driving Simulation · 1Efficient VIO · 1Egocentric data · 1Egocentric video · 1Embodied Agents · 1Embodied policy · 1Empowerment MPC · 1Endoscopic navigation · 1Endovascular robotics · 1Event-camera VO · 1Exoskeletons · 1Failure recovery · 1Federated 3D Perception · 1Fisheye detection · 1Floorplan SLAM · 1Geometry-aware manipulation · 1Grasping · 1Grounding · 1Humanoid Control · 1Humanoid RL · 1Humanoid Tool Use · 1Humanoid control · 1Humanoid loco-manipulation · 1Humanoid safety · 1Imitation Learning · 1JEPA World Models · 1Lab automation · 1Laboratory Manipulation · 1Latent World Models · 1Legged Manipulation · 1LiDAR Security · 1Lifelong RL · 1Locomotion · 1MAPF · 1MAV guidance · 1Manipulation LfD · 1Manipulation Safety · 1Marine simulation · 1Memory-centric VLA · 1Miniature Locomotion · 1Mining Perception · 1Mobile manipulation · 1Model-based RL · 1Monocular SSC · 1Motion Planning · 1Multi-Agent RL · 1Multi-Robot Exploration · 1Multi-Robot Manipulation · 1Multi-UAV Exploration · 1Multi-UAV State Estimation · 1Multi-agent RL · 1Multi-agent motion planning · 1Multi-camera odometry · 1Multi-fidelity RL · 1Multi-robot manipulation · 1Multi-skill Manipulation · 1Multi-view perception · 1Multimodal simulation · 1Navigation & Mapping · 1Object Pose Estimation · 1Object Search · 1Panoramic 3D Detection · 1Panoramic segmentation · 1Pedestrian tracking · 1PersonNav · 1Planning · 1Policy · 1Policy Optimization · 1Preference Learning · 1Quadruped locomotion · 1Real-time VLA · 1Real-to-sim · 1Representation alignment · 1Robot-Use Agents · 1Runtime Verification · 1Safe Generative Planning · 1Safe RL-MPC · 1Safe VLA · 1Safety monitoring · 1Safety-critical control · 1Sim-to-real Navigation · 1Sim-to-real manipulation · 1Simulation geometry · 1Skill Discovery · 1Social navigation · 1Soft Pneumatic Actuators · 1Soft manipulation · 1Sonar SLAM · 1Streaming 3D reconstruction · 1Submovement decomposition · 1Surgical robotics · 1Tactile · 1Tactile Exploration · 1Tactile VLA · 1Temporal manipulation · 1Terrain traversability · 1Text-to-3D Policy · 1Tool Co-Design · 1Trajectory Correction · 1Traversability · 1UAV Navigation · 1UAV Swarms · 1UAV traffic monitoring · 1Underwater Robotics · 1Underwater active mapping · 1Underwater perception · 1Unified world-action model · 1VLA Conditioning · 1VLA Defense · 1VLA Robustness · 1VLA evaluation · 1VLA manipulation · 1VLA robustness · 1Video world models · 1Visual Generalization · 1Visual Grounding · 1Visual Reward Models · 1Visuomotor Fusion · 1Visuotactile manipulation · 1World Models · 1World-Action Models · 1
LevelAnyBeginnerIntermediateExpert
arXiv
Multi-Agent RLIntermediate

Permutation Robustness Is Not Enough: Action Collapse in Multi-Agent Transformer Policies

Amit Thakur, Mukesh Singhal

This preprint examines transformer policies for cooperative multi-robot navigation and shows that measuring permutation equivariance alone is misleading: low error can occur because all agents output the same action, not because the policy is robust. In a PettingZoo/MPE cooperative navigation task, a PPO baseline with learned agent ID/position embeddings is diverse but order-sensitive, while equivariance regularization reduces order error but can collapse actions, especially when moving from 3 to 4 agents. The work argues for reporting task return, permutation consistency, and action diversity together.

Read the full paper (PDF)
arXiv
ExoskeletonsExpert

Subject-Specific Predictive Musculoskeletal Simulations of Lower-Limb Exoskeleton Assistance: Metabolic and Biomechanical Effects of Joint Assistance Strategies

Neethan Ratnakumar, Mariya Tohfafarosha, Xianlian Zhou

Preprint. Researchers built subject-specific musculoskeletal walking simulations with idealized hip, knee, and ankle actuators and tested all single- and multi-joint combinations at 25 and 50 Nm. At 50 Nm, three-joint hip–knee–ankle assistance cut predicted metabolic cost by 48.50% on average, while hip–ankle assistance captured most of that benefit at 44.60%, supporting simpler two-actuator designs. The results provide simulation-derived guidance for exoskeleton developers but still need hardware validation.

Read the full paper (PDF)
arXiv
Contact-Rich ManipulationIntermediate

Learning Reflexive Behavior for Contact-Rich Manipulation

Quan Nguyen, Yunho Kim, Joonho Lee

This preprint introduces a learned proprioceptive reflex policy that sits between task-level commands and joint control. Trained in simulation on spring, plane, and rail interaction primitives, it keeps contact forces near thresholds when commands or geometry are imperfect; on real peg insertion with 0.02 mm clearance, it raises success from at most 22% to 36–58% and uses less than half the mean estimated contact force of IK+PD. The result matters for assembly, humanoid lifting, and surface following as a reusable low-level execution layer under planners, learned policies, and teleoperation.

Read the full paper (PDF)
arXiv
Visuotactile manipulationExpert

Equivariant Visual-Tactile Diffusion Policy for Contact-Rich Manipulation

Lik Hang Kenny Wong, Yiyao Ma, Xiu-Shen Wei, Zelong Tan, Zhuheng Song, Dongsheng Xie, Kai Chen, Qi Dou

This preprint introduces VISTA, a diffusion-policy architecture that fuses wrist camera and tactile images into a rotation-equivariant spherical representation for contact-rich manipulation. In simulation with 50 demonstrations per task it reaches 80.9% mean success, outperforming an equivariant tactile baseline by 15.9 points, and on three real-robot tasks it succeeds in 7.33/10 trials versus 1.33/10 for that baseline. The design could reduce the demonstration data needed for tight-insertion and surface-contact tasks by encoding workspace rotation symmetry into visuotactile perception.

Imitation Learning
Read the full paper (PDF)
arXiv
Endovascular roboticsExpert

Autonomous Robotic Navigation for Endovascular Brain-Computer Interface Access

Harry Robertshaw, Weijie Qi, Nikola Fischer, Alejandro Granados, Thomas C. Booth, Sam E. John

This arXiv preprint reports the first demonstration of autonomous robotic navigation through cerebral veins for endovascular brain-computer interface access. Two Soft Actor-Critic tasks were trained in simulation and tested on 250 simulated and five physical fluoroscopy-guided runs per condition; overall in vitro success was 70% (14/20), but hold-out Task A fell to 40–42% and the failure predictor lost calibration in physical tests. The work establishes feasibility for reducing operator burden in BCI delivery, though substantial validation remains before clinical use.

Read the full paper (PDF)
arXiv
Lifelong RLExpert

How to Find and Reuse Policies for Continuous Adaptation in Lifelong Reinforcement Learning

Saptarshi Nath, Inish M. D'Souza, Antonio Carta, Soheil Kolouri, Andrea Soltoggio

In this preprint, researchers introduce Adaptive Mask Selection and Composition (AMSC), a lifelong reinforcement learning method that uses online task-similarity embeddings to choose and weight previously learned policy masks when adapting to new tasks. On CT28 and MiniGrid benchmarks, it reaches AUC/forward transfer of 0.94/0.94 and 0.73/0.54 with zero measured forgetting, exceeding modular baselines. The approach matters because it enables selective policy reuse instead of reusing an entire history, reducing interference and helping agents accumulate skills without destructive forgetting.

RL
Read the full paper (PDF)
arXiv
Sim-to-Real ManipulationIntermediate

Skill2Real: Agentic Skill Learning for Zero-Shot Sim-to-Real Robot Manipulation

Xincheng He, Siyu Ma, Chang Yu, Yunuo Chen, Yanjia Huang, Ying Nian Wu, Yin Yang, Chenfanfu Jiang

Skill2Real is a preprint describing an agentic framework that learns reusable robot manipulation skills in simulation and transfers them zero-shot to real hardware through a shared code-based policy interface. On a physical UR5e, it raises mean task completion from 27.50% with no learned skills to 78.75% across four tasks, and on unseen long-horizon LIBERO-Pro Long tasks it raises GPT-6 Astra success from 2.0% to 56.3% using frozen skills. This matters because it shifts sim-to-real learning from adapting low-level representations to acquiring validated, composable task knowledge without real-robot fine-tuning.

Read the full paper (PDF)
arXiv
ManipulationIntermediate

Proprioceptive Sketches as Long-Horizon Intent for Generative Action Policies

Fangyuan Wang, Songhao Huang, Haoxiang Sun, Shipeng Lyu, Chengyang He, Anqing Duan, Peng Zhou, David Navarro-Alarcon

The paper presents Proprioceptive Action Models (PAM), a preprint that adds a compact, timing-free B-spline sketch of the robot’s remaining joint-space path to short-horizon action generation inside one diffusion transformer. Across real-world dual-arm tasks, PAM-DP raises success from 47.5% to 75.0% and also improves Push-T and LIBERO-Long. This offers long-horizon intent without costly language plans or video forecasting.

Read the full paper (PDF)
arXiv
VLA DefenseExpert

Detect and Suppress: A Mechanistic Defense against Adversarial Patches in VLA Models

Yukiya Horiba, Koshiro Aoki, Shunsuke Yasuki, Bum Jun Kim, Taiki Miyanishi

Preprint. The authors build an inference-time defense for vision-language-action (VLA) policies against adversarial image patches by training a sparse autoencoder, identifying an internal feature that tracks patch presence, and suppressing that feature only when a linear probe detects an attack. On LIBERO-10 under intermittent UADA attacks, conditional suppression raises success rate, e.g. from 26.0% to 32.4% for π0.5 at n=10, while avoiding the large benign-condition drop caused by continuous suppression. This points to lightweight, fine-tuning-free robustness layers for deployed robot policies.

VLA
Read the full paper (PDF)
arXiv
Real-to-simExpert

LiteReality-Agent: An Agentic System for Interactable 3D Indoor Scene Reconstruction

Zhening Huang, Yueyan Li, Johnathan Chiu, Xiaoyang Lyu, Matt Zhou, Yuxin Yao, Joan Lasenby, Shangzhe Wu

This preprint introduces LiteReality-Agent, an agentic system that converts consumer RGB-D room scans into editable, articulated and MuJoCo-ready 3D scenes by having a coding agent iteratively maintain an executable Python scene program. On ten rooms and 620 capture views it reports PSNR 14.13 dB and depth MAE 10.5 cm, with zero detected furniture intersections, beating GPT-6 and Fable 5.1 baselines. The public code and capture app give robotics teams a practical real-to-sim authoring foundation.

Read the full paper (PDF)
arXiv
Multi-agent RLExpert

Test-time Multi-agent Coordination by Decomposed Value Gradient Flow

Dongsu Lee, Haoran Xu, Amy Zhang

This preprint introduces SCOUT, an offline multi-agent reinforcement learning framework that pairs a per-agent flow-matching behavior prior with a decomposed value function and applies test-time Stein variational refinement to action candidates. Across SMACv1 and MA-MuJoCo, it reports the best average scores among all listed baselines (SMACv1 average reward 16.1 vs 14.9 for MAC-Flow; MA-MuJoCo normalized score 82.12 vs 76.37 for MAC-Flow) and improves on all reported offline-to-online transitions. The result matters because it shifts the coordination/value trade-off from a fixed training-time regularization to an adjustable inference-time compute budget.

RL
Foundation
Read the full paper (PDF)
arXiv
World Action ModelsIntermediate

Completion Aware Guidance for World Action Models

Seungyeon Kim, Junhoo Lee, Baekseung Kim, Minkyu Kim, Nojun Kwak

This preprint introduces Completion Aware Guidance (CAG), a training-free sampling method that steers world action models to generate task-completing video-action chunks instead of plausible but incomplete continuations. On a RoboTwin 2.0 subset it lifts average success from 64.4% to 70.0%, and on DreamZero zero-shot simulation from 69% to 75%, while cutting task-incomplete imagination from 79% to 40%. It matters because it improves existing WAMs without retraining, reducing a distinct failure mode in short-horizon robot control.

Read the full paper (PDF)
arXiv
Streaming 3D reconstructionIntermediate

CLoSeR: Closing the Loop for Long-Context Streaming Reconstruction

Moyang Li, Zihan Zhu, Wei Zhang, Marc Pollefeys, Daniel Barath

This preprint presents CLoSeR, a loop-closure system built on the streaming feedforward reconstructor LoGeR for long monocular videos. It detects revisits with global image descriptors, re-runs the backbone on loop-conditioned windows, and optimizes all poses on SE(3), cutting average trajectory error substantially on kilometer-scale benchmarks (e.g., 11.15 m vs 36.16 m for LoGeR on VBR). Because the backbone already has consistent global scale, this avoids the fragile point-cloud alignment and higher-dimensional pose graph optimization used by earlier submap methods.

Read the full paper (PDF)
arXiv
Autonomous DrivingBeginner

End-to-End Learning vs. Modular Architectures: Comparative Insights into Autonomous Driving Systems

Kartik B. Kapse

This preprint introduces a Four-Dimensional Architecture Selection Framework that compares modular, end-to-end imitation learning, end-to-end reinforcement learning, and hybrid autonomous driving system designs. Tested against ten published systems from 1989 to 2023, the framework correctly identified the deployed architecture in seven of ten cases (four exact matches and three partial matches). It offers a structured decision-support tool for teams choosing an autonomous driving architecture, though it does not itself build or test a new driving system.

Read the full paper (PDF)
arXiv
Humanoid safetyIntermediate

Continue, Abort, or Fall: Viability-Aware Policy Selection (VAPS) for Safe Humanoid Acrobatics

Siwei Ju, Lu Liu, Jan Peters, Oleg Arenz

Preprint. VAPS is a safety layer for humanoid acrobatic robots that adds two backup policies—an abort policy that lands on its feet and a protective-fall policy—and selects among them plus the nominal tracking policy using short-horizon viability predictors. In simulation on a LimX Oli, it cut head-contact rate from 15.5% to 1.1% while keeping motion success nearly unchanged (79.2% vs 80.7%), and on the physical robot the full hierarchy achieved its intended objective in 15 of 18 side-flip trials. The work matters because it directly addresses the main cost of deploying dynamic humanoids: hardware damage from failed maneuvers.

Humanoid
Read the full paper (PDF)
arXiv
Pedestrian trackingIntermediate

Beyond Leaderboard Scores: A Deployment-Focused Protocol for Interpretable Tracking Evaluation in Pedestrian-Centric Environments

Dominik Wojcikiewicz, Diego Paez-Granados

This preprint presents a deployment-focused evaluation protocol for 3D pedestrian multi-object tracking that holds detections fixed and reports trajectory-level success rates for initialization, detector-gap continuation, identity recovery, close-neighbor association, and load-dependent runtime on a Jetson Orin. On JRDB, non-GT-assisted trackers score only 24.26–29.67% HOTA; after a 1.0 s detection gap, no non-GT tracker maintains correct same-ID spatial output in more than 45.4% of cases, and most fall below 1%. The protocol shows aggregate leaderboard scores hide large behavioral differences that matter for real-time robots among pedestrians.

Read the full paper (PDF)
arXiv
Robot-Use AgentsIntermediate

OpenRUA: Robot-Use Agents Are Zero-Shot Visuomotor Policies

Zhaoyang Chu, Earl T. Barr, Claire Le Goues, Peter O'Hearn, Mark Harman, Federica Sarro, He Ye

OpenRUA is a preprint describing a zero-abstraction harness that lets off-the-shelf coding agents control a robot through its native ROS 2 terminal interface, without bespoke primitives or learned policies. In simulation, Claude Code powered by Claude Opus 5 reached 99.0% success on CaP-Bench and 87.0% on LIBERO-PRO, and on RoboCasa365 it held unseen composite task performance at 28.1% while baselines collapsed. This matters because it suggests frontier coding agents plus a mature robot middleware stack may be a practical zero-shot alternative to training VLAs or hand-engineering agent wrappers.

ROS 2
Read the full paper (PDF)
arXiv
Navigation & MappingIntermediate

GlassGuard: Verified Glass Plane Mapping for Robot Navigation

Hanwen Guo, Zhengzhi Lin, Yusen Xie, Ji Zhang

GlassGuard is a preprint system that reconstructs planar architectural glass for LiDAR navigation by combining a pruned SAM3 vision model, LiDAR-derived structural pillars, depth-free orientation checks, and persistent map management. On nine real-world scenes and 2.1 km of traversal, its panoramic version covers 85.2% of glass with 15.3 false voxels per frame, while the pinhole version covers 82.1% versus at most 61.0% for baselines and carries 5–17× fewer false voxels per frame. It matters because it blocks collision paths through glass without contaminating traversable space, a dual requirement for safe robots in glass-walled buildings.

LiDAR
Read the full paper (PDF)
arXiv
Imitation LearningExpert

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

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

This arXiv preprint introduces EF-GAIfO, a feasibility-aware imitation-from-observation method that learns robot policies from state-only human demonstrations by scoring each demonstrated transition against the robot's own accumulated experience. On a Unitree Go2 quadruped, policies learned with EF-GAIfO reached 70–80% real-world object-grasping success versus 10–30% for WGAIfO and GAIfO baselines. The approach matters because it removes the need for action-labeled demonstrations or explicit robot dynamics models when learning from human motion that the robot may not be able to reproduce.

Read the full paper (PDF)
arXiv
World ModelsExpert

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

Jiahui Lei, Qianqian Wang, Trevor Darrell, Angjoo Kanazawa

This preprint, World Motion Models (WMMs), casts dynamic 3D scenes as collections of sparse SE(3) pose trajectories and trains a single flow-matching transformer to answer arbitrary known/unknown conditioning queries. Across separate checkpoints it reports 87.2% macro success on Language Table (vs 80.0% for LCB, the best prior), a 90.0% success on B2B, and faster/higher-success humanoid retargeting than offline PHUMA. It matters because the same architecture can serve as policy, forward dynamics model, planner, and cross-embodiment retargeter.

Read the full paper (PDF)

Topic pages

Manipulation · 27VLA · 21Perception · 12General robotics · 11Navigation · 10Humanoid · 9Robot learning · 8Safety · 5World Action Models · 5Simulation · 3Aerial manipulation · 3Dexterous manipulation · 3Sim-to-Real Manipulation · 3Social Navigation · 3

Papers come from arXiv robotics feeds. Where our summary is not written yet, you see the opening lines of the abstract.