Perfect discrimination can still make certified admission impossible
Telling two states apart and certifying an action need different sensing designs — and different audits.
cs.RO
Robotics
Roughly includes material in ACM Subject Class I.2.9.
sort pith recommended most recent
TANGO navigates humanoids through clutter with language and whole-body actions
Directly predicts 29-DoF commands from RGB and text, achieving zero-shot sim-to-real transfer on a Unitree G1 robot.
DeCAL is a vision-language-action model for dexterous robot manipulation that adaptively…
DeCAL unifies understanding, imagination, and action generation with contact-aware tactile fusion to achieve state-of-the-art dexterous…
Interval reachability ensures safe sampling MPC without pre-computation
Reduces crashes by over 99% in simulation and runs at 60 Hz on a real racecar.
· “Online, Reachability-Aware, Sampling-Based Motion Planning”
Observation gating revises learned occupancy for better active mapping
Controlled experiments show ground-truth occupancy speeds coverage but not final coverage; a filter recovers targeted failures without…
· “Rethinking Learned Occupancy in Autonomous Active Mapping with Observation-Gated Filtering”
Voronoi spreading speeds recovery in multi-vehicle tracking
Field tests show a simple particle-spreading heuristic reduces peak error after communication dropouts without harming nominal performance.
· “A Distributed Consensus Particle Filter for Target Tracking using Autonomous Surface Vessels”
DYAD is a multimodal dataset of 20 human-human gearbox assembly sessions
DYAD records 851 co-located assistance events linking requests, verbal and physical help, triggers, and outcomes during gearbox assembly.
· “DYAD: A Multimodal Dataset of Co-Located Human Assistance”
SphereOcc tops 3D occupancy from 360° images by 1.7 points
A new dataset and two geometry-aware modules bridge the angular-to-Cartesian gap for autonomous driving perception.
· “Spheriverse: 3D Scene Understanding from Spherical Observations in the Wild”
LLM agents unify all stages of autonomous driving scenario tests
PlannerForge outperforms specialized tools on generation, selection, modification, and enhancement without fine-tuning.
· “PlannerForge: LLM Agents for Scenario-Based Testing of Motion Planners in Autonomous Driving”
Contact-aware GNN steers tensegrity robots over obstacles
A learned dynamics model augmented with differentiable contact detection enables a hybrid MPPI controller to navigate ramps, narrow…
· “Model Predictive Control of Tensegrity Robots via Contact-Aware Graph Neural Dynamics Model”
Robot motion carries secret messages decodable from video
New method encodes arbitrary bits as subtle noise in robot behavior, allowing auditors with a shared key to recover intent from remote…
Actuated cameras triple usable workspace on humanoid
New visible-reachable metric cuts dual-target grasping time by 17% and energy by 19%.
· “Visible-Reachable Workspace for Perception-Aware Humanoid Design”
Factored attribute retrieval tops scene memory baselines by 15 points
FRAME decomposes language queries and learns per-attribute probes, enabling fast and accurate multi-attribute object recall from fixed…
· “FRAME: Factored Retrieval via Attribute Readouts for Object-Centric Scene Memory”
Boost sample efficiency with alternating state-value targets
CAST uses alternating planner and learned-policy transitions to train a state-value critic, boosting sample efficiency on 14…
One minute to a playable 3D world from a single video
A feed-forward framework parses objects and generates complete textured meshes without iterative optimization or manual prompts.
· “FIRE3D: Feed-forward Interactive 3D Scene Reconstruction Within A Minute”
One IMU localizes puncture in soft actuator at 96% accuracy
Chamber-by-chamber perturbation and an autoencoder let untethered soft robots detect damage without extra hardware.
· “Real-time Puncture Detection and Recovery for Pneumatic Soft Actuators”
Zero-collision MARL via LiDAR safety filter
A safety-decoupled framework guarantees collision-free cooperative navigation under time-varying topologies.
· “Graph-Based Safe Reinforcement Learning for Multi-Agent Systems with Time-Varying Topology”
Takes 50x larger steps through stiff contact
Keeps gradients reliable at 0.1 s steps: 8,192 worlds on one GPU, control tasks MJX cannot solve.
· “Ostrich: Taking Large Strides Through Stiff Contact in Differentiable Dynamics”
Teleop robot triples time and fragments hand motion
Same instrument, same eye model, same targets: swapping in a robot changed every movement measure.
One demonstration teaches a robot to manipulate objects via relative motion
FOCI Policy extracts short interaction segments and models object-relative trajectories, achieving strong data efficiency and…
· “FOCI Policy: Focus on Object-Centric Interactions for Relational Manipulation Policies”
Footprint clearance lifts navigation success from 42% to 70% in simulated tests
Replacing sensor range with distance to the robot body in LiDAR features yields a 28-point gain, though seed selection limits the result.
· “DCLP++: Learning to Navigate with Footprint Clearance and Relative Motion”
Bayesian dialogue manager beats flat models
Across three offline datasets, Bayesian neural nets outperform deterministic variants; an online study shows hierarchical explanations…
Frozen tactile models jump from 6.86% to 87.09% on unseen sensors
With just 10% labeled contacts on the new sensor, no retraining or paired calibration is needed.
· “BIFTA: Brain-Inspired Few-Shot Tactile Adaptation for Unknown Sensors”
Hybrid RL lets robots build spanning arches adaptively
A novel extension of soft actor-critic handles mixed actions and graph states, achieving real-robot closed-loop assembly.
· “Learning to build covering structures with continuous adjustments”
SUN: multiplicative goal score beats additive baselines in exploration
A product of successor value and pseudocount jointly scores novelty and reachability, outperforming tuned additive methods with fewer…
Chunk-spanning semantics lift robot success up to 7 points
Occupancy-weighted stage descriptions beat current-step labels on LIBERO, RoboTwin, and MolmoSpaces; one variant declines.
· “CASD: Chunk-Aligned Semantic Distillation for Multi-StageRobot Manipulation”
Multi-fisheye VINS cuts trajectory error sixfold
Omnidirectional MFVINS with learned depth averaged 0.21 m error vs 1.39 m for monocular VINS and 2.61 m for MCVIO.
· “MFVINS: Multiple Fisheye Camera-Based Visual Inertial System”
Gaussian mixture model reveals reliability of VLM traversability estimates
New method gives robots a statistical way to know when their terrain understanding is unreliable, enabling safer autonomous navigation in…
Sector-guided robot explorer triples rugged-hill coverage
Coarse region wedges set the route, fine viewpoints fill the gaps, and rough terrain gets mapped in about half the time.
· “TASG-Explore: Traversability-Aware Sector-Guided Exploration for Ground Robot on Uneven Terrain”
One policy tracks whole-body motion over 37-cm obstacles
PGMT uses motion-conditioned terrain glimpses to adjust footholds and posture, enabling a Unitree G1 to handle stairs, rough ground, and…
· “PGMT: Perceptive General Motion Tracking for Humanoid Robots”
Picking each spin by uncertainty lifts in-hand reconstruction to 86%
A robot hand that picks its next view from uncertainty beats fixed spin schedules on a 30-second budget.
· “AURORA: Active Uncertainty-Driven Re-Orientation for In-Hand Reconstruction”
Soft robot arm plays 12-bounce drum rolls with two tuned motion tricks
Tap-Pull and Micro-Pulse, tuned by Bayesian optimization, turn passive elasticity into sustained, uniform drum sounds.
· “Multi-bounce Drum Roll with Optimized Active Tricks to Leverage Soft Embodiment”
Graph memory lets robots dodge hidden hazards
By storing object relationships even after they leave view, SafeMem lifts safe success rate from 0.21 to 0.59.
· “Safe Task Planning with Long-Term Graph Memory for Embodied Agents”
Spatial anchors boost zero-shot aerial navigation by 39%
AirAnchor bridges local and global space via shared object anchors, outperforming all zero-shot baselines on AerialVLN-S.
Attention pooling on early CNN features lifts robot success 36-41%
FocusPool learns task-progressive masks from intermediate layers, needing half the training data of baselines.
· “Localized Visual Feature Aggregation via Focus Pooling for Visuomotor Policies”
Generalized spanning trees cut joint motion for arm coverage
By selecting one IK solution per cell, the method reduces joint movement and avoids revisiting sub-cells.
· “Coverage Path Planning for Redundant Manipulators using Generalized Spanning Trees”
Depth‑free floorplan localization matches depth models
Enforcing verticality and coplanarity on RGB wireframes achieves 88% sequential success at 0.1 m on Gibson, outperforming depth‑based F3Loc.
· “GALoc: Gravity Aligned Wireframes for Depth-Free Monocular Floorplan Localization”
RoboCousin turns photos into robot-training playgrounds
Automatically generates millions of bimanual manipulation demonstrations from user images, improving real-world transfer.
· “RoboCousin: Build Your Own Simulation Playground for Robust Bimanual Robotic Manipulation”
Late-fusion tracker fuses five sensors to see race cars at 96 meters
Real-world tests show camera-LiDAR-radar combinaton beats any pair in speed, range, and accuracy, enabling safe high-speed overtaking.
· “A Multi-Modal Perception Pipeline for Object Detection and Tracking in Autonomous Racing”
Evolving rooms break lifelong navigation memory
New benchmark shows that reusing stale scene representations cuts success rate by up to half in changing environments.
· “EvoNav-Bench: Benchmarking Lifelong Navigation in Evolving Environments”
One env variable hijacks robots
Attackers replace telemetry before encryption; a Franka arm was hijacked 100% of the time while verifiers saw normal data.
· “Seeing is Not Believing: Breaking the Physical-to-Digital Trust Boundary in Robotics”
Clean scenes can't rank physics inference in vision encoders
CALIPER test shows varying camera and clutter is needed to separate models that truly infer mass and friction from those that don't.
· “CALIPER: Clean Scenes Cannot Rank Physical Inference in Pretrained Visual Representations”
Multi-view 2D waypoints yield 3D robot guidance for unseen tasks
Predicting consistent 2D points across cameras and triangulating them gives robots explicit 3D motion cues for generalisation.
· “3DWay: Generalizing Robot Manipulation via 3D Consistent Waypoints”
LLM iterative pipeline designs real soft robots from text
A closed-loop framework with simulation and critique achieves 96.2% validity and produces three working grippers.
· “Bridging Language and Physics: Automated Design of Continuum Robots with Large Language Models”
Memory-tool loop lifts driving VLM accuracy by 7.7 points
Short-term scene memory and long-term experience combine to beat larger models on driving benchmarks.
Clip-on fingertip sensor reads forces under 0.5 N
Fiber-optic clip captures contact force and texture vibrations for robot teaching, wet or dry.
· “TacClip: a clip-on sensor measures dynamic contact forces without covering the fingerpads”
Separating hunches from evidence lifts drone search to 21.61%
A dual-layer spatial memory turns flickering AI sightings into steady search guidance, with best marks on the UAV-ON benchmark.
· “Dual-Layer Semantic-Spatial Belief Mapping for Aerial Object Goal Navigation”
One factorized posterior guides long-horizon robot navigation
OmniNav keeps memory, target belief, and reachability in sync, lifting real pick-and-place success from 53% to 72%.
· “OmniNav: Robust Long-Horizon Target Navigation in Dynamic Environments”
Letting an alert AI say 'observe' first improves crash warnings
A tri-action policy raises dashcam AUROC from 0.610 to 0.689 and catches 89% of takeovers within five seconds.
· “Observe Before You Alert: Adaptive Driver Alerting with Vision-Language Models”
Choosing which failure to fix wins 10 of 10 settings
Grouping failed episodes into recurring modes and targeting each new demonstration yields the largest edge at the smallest budget.
· “DISEIL: Demonstration Distillation for Sample-Efficient Imitation Learning”
Sjngle-precision GPU achieves millimeter orbital accuracy
mjorbit couples MuJoCo with Encke's method in an orbit-following frame, matching Basilisk at 1e-7 m while running 1000× faster.
SPOT's wide stable view cuts humanoid teleop times by up to a third
Decoupling the operator's gaze from robot motion speeds search, recovery, and bimanual tasks in VR teleoperation.
· “SPOT: Spatial Perception-Oriented Long-Horizon Humanoid Teleoperation”
EMP model forecasts falls 12% better on non-periodic data
Leveraging planned motion trajectories as exogenous variables, it outperforms state-of-the-art on real and simulated aperiodic benchmarks
· “A Multimodal Label Forecasting Method for Aperiodic Visuo-Motor Time Series”
Force-touch controller cuts contact loss 99% in robot wiping tasks
M3-Tele combines force and tactile feedback to produce cleaner, learnable demonstrations for mobile manipulation policy learning.
Scene graph haptic cuts needle error 14% in robot eye surgery
Real-time semantic map drives vibra feedback; users align early without time penalty
Sparse anchors plus diffusion keep image meaning at 0.039 bits per pixel
CLIP picks the few latent cells that matter; a tuned diffusion model fills the rest, even on images never seen before.
· “Foundation Models for Generalizable Semantic and Goal-Oriented Communication”
Sending 32 of 512 image tokens keeps robot success near 95%
The instruction knows what matters: sending 32 of 512 image tokens costs just 1.5 points of task success.
· “ComVLA: Communication-Aware Split Inference for VLA Models in 6G-Connected Robotics”
Simulated touch fills in what human videos miss for robot dexterity
DEX-X learns visual-tactile manipulation from video and transfers zero-shot to a real hand-arm system, achieving 93% cube-picking success.
· “Dex-X: Learning Visual-Tactile Dexterous Manipulation From Human Videos with Simulated Interaction”
Decentralized shield lets robots avoid collisions without talking
Predictive safety layer and a deadlock-breaking protocol allow pre-trained multi-agent policies to adapt safely to unseen obstacles.
· “Decentralized Safe Multi-Agent Reinforcement Learning via Predictive Shielding”
Aligned demos steer a frozen robot policy to 97.7% on LIBERO
Retrieved demonstration snippets give a frozen text-only policy a 19-point edge on RoboTwin 2.0.
Flying humanoid walks a ceiling via thrust-rate control
A 0.15 m four-step ceiling walk shows that commanding thrust derivatives instead of thrust kills the spikes that break contact.
Zero-shot sim-to-real assembly: 93% success via proprioception anchoring
PACE encoder learns domain-invariant features by predicting joint transitions, enabling direct policy deployment without real data.
· “Zero-Shot Sim-to-Real Contact-Rich Assembly via Proprioception-Anchored Cross-Modal Pretraining”
One RGB-D video yields a soft object's hidden physics
PhysReal's hybrid material model predicts how toys and cloth deform under new pushes better than four baselines.
· “PhysReal: Learning Real-World Deformable Object Physics via Hybrid Constitutive Modeling”
Two CAD priors slash LiDAR-camera calibration error by 96%
Fixing hole radius and rectangle layout eliminates need for dense LiDAR, extending usable standoff range 1.5×.
· “P²Calib: Utilizing Pattern Priors for LiDAR-Camera Extrinsic Calibration”
Road maps forecast 100–150 m ahead with 7.83 m error
First demonstration that learning-based methods can plausibly extend HD maps beyond sensor range on simple roads, offering a path toward…
Flow matching beats token-by-token policies on unseen robot tasks
Chunked continuous actions from one demonstration rival per-task fine-tuning in simulation and on real bimanual arms.
· “ContextFlow: In-Context Flow Matching for Robot Manipulation”
Perfect world maps barely boost robot coverage
Diagnosis over 25 starts finds occupancy fixes mainly speed exploration; a new filter rescues two hard cases.
· “Diagnosing and Dynamically Filtering Occupancy World Models for Active Mapping”
Robotaxi routing can improve both monitoring and speed
Simulation shows a win-win when traffic coverage is added to route planning with the right weighting.
· “RoboSense: Leveraging Robotaxi Fleets as Drive-by Sensors for Urban Traffic Monitoring”
Neural OCT field enables 43× faster queries
A hybrid neural representation turns discrete OCT into a continuous tissue field, speeding robotic path planning and sparse reconstruction.
· “OCTN: Neural OCT Representations for Robot-Guided Precision Intervention”
Spherical feet give underwater quadruped attitude control
A simplified model and off-the-shelf parts let a four-legged robot turn and tilt without propellers.
· “Design and Attitude Control of an Underwater Quadruped Robot”
One policy to dribble, trap, and shoot on a humanoid
SkillX unifies three soccer skills in a single neural network and transfers from simulation to a real robot.
· “SkillX: Unified Multi-Skill Policy Learning for Humanoid Soccer”
Pipeline grounds robot scene descriptions with real geography
Augmenting VLM outputs with OpenStreetMap and sensor data yields building F1=0.83 in outdoor tests.
· “Physico-Geospatial Grounded Scene Interpretation for Mobile Robotics”
No map, no problem: a compact memory guides exploration
Diffusion planner with a 512-dim recurrent state covers unseen spaces in simulation and flies an Elios 3.
· “CAVEAT: Recurrent Multimodal Diffusion Planning for Mapless Aerial Exploration”
Explicit 3D geometry lifts diffusion policy to 96.67% success
Outperforms vision baseline by 8.67 points under combined position and viewpoint shifts.
Three-stage training unifies humanoid locomotion and object manipulation
A context-conditioned discrete skill prior enables reusable interaction skills without task-specific reward engineering, achieving 88%…
· “Unifying Physics-Based Humanoid Interaction with a Context-Conditioned Interaction Prior”
3D mapping runs up to 4.07x faster by fusing flat pixels as one ray
Walls and floors stop being re-fused every frame; mesh error stays under 0.19 cm while TSDF cost drops.
Hierarchical learning achieves 0.1mm cylinder assembly in 48 steps
Simulated in 500 episodes, transferred directly to a real robot with 100% success over 20 trials.
Drone plus camera gimbal exactly mimics fixed-wing flight
A control proof and simulation show a camera drone can replace a costly fixed-wing plane for training and testing.
· “Dynamic System Emulation: Fixed Wing Dynamics on a Multicopter”
Observable history unlocks recovery for frozen VLA policies
A learned verifier uses past images, actions, and states to detect failures and issue stage-aware fix prompts, boosting success without…
People can't tell human from AI in short robot teleoperation clips
A 50-person study finds no reliable difference in perceived agency between human-operated and AI-generated robot behavior.
