Robotics
Research on Robotics: explanations, evidence, limitations and connections between papers.
21 results
Training robot policies to act after they move the camera
ActiveAlign adapts robot training to camera-moved views, improving task success in perturbed-start simulation and physical occlusion trials.
A smoother keeps violating samples in the trajectory update
A trajectory method keeps smoothed safety estimates below exact values and retains violation gradients in robot simulations.
Robot touch training improves control with and without touch
Spark-0 uses tactile pretraining to improve vision-only robot manipulation scores on the tested Ego Humanoid tasks.
A driving test’s comfort score can reward staying still
A corrected simulator comfort score stops tracking time spent stopped across eight models, changing how driving comfort is assessed.
Robot alarms can predict failure yet still raise task costs
RESCUE tests robot alarms with stop-and-reset; evaluated monitors did not lower task-completion cost versus no intervention on primary LIBERO-Plus tasks.
A mapless depth-image query speeds reactive drone collision checks
A GPU method queries depth images directly, reducing collision-check latency in a tested 400-query drone-planning setting.
One robot model learns to predict scenes and control varied robots
GreenWAM’s robot-specialized starting point improves RoboTwin task success over its parent under a fixed adaptation recipe.
Match predictive evidence to the robot decision it serves
A robot facing a moving object needs recent motion to choose an action. A driving planner needs evidence to compare possible paths. After execution, a model may need evidence about what an action actually caused. Each decision asks something different of prediction. These studies show predictive states and visual history feeding action generation, a forecast informing path selection, and verified replays improving discrimination between successful and failed actions. Their evaluations support those particular roles in their tested settings; a better prediction measure alone does not establish better control.
A driving planner organizes candidate paths by motion
ManeuverFlow groups driving paths by motion and uses shared future-scene evidence to score and refine them.
Robot models improve when actions explain intended effects
Arc adds causal traces to existing robot demonstrations and improves control across tested simulation and hardware tasks.
World-model rehearsal helps robots retain skills while learning
D2R helps robots learn sequential tasks while retaining earlier skills through regenerated world-model rehearsal.
Reflex helps a humanoid catch thrown boxes with moving-camera vision
Reflex separates body control, box-motion prediction, and vision so a humanoid can catch thrown boxes, with long-range limits.
Robots can fail during learning, planning, or instruction following
A robot can falter after learning another task, choose poorly among driving paths, or respond differently to two phrasings of a request. These failures arise at different decisions, so the evidence that might help enters at different points. The studies suggest rehearsing earlier tasks during adaptation, consulting predicted scenes while selecting a path, or rewriting an instruction before action. Rehearsal reduces stored trajectory memory while using model computation; the driving planner shares forecast work across paths. Each reported gain belongs to its own test.
GLIO2 completes a route with weak LiDAR geometry
GLIO2 jointly updates LiDAR, inertial, and satellite measurements, completing a tested degenerate bridge route where LiDAR systems diverged.
Human motion helps humanoid robots follow instructions
VioLA uses shared human and robot motion targets to train humanoid control, with strong results in one laboratory evaluation.
Rewarding diverse robot successes improved tested shift averages
DRIVE rewards diverse successful robot behavior; simulation and two physical tasks show higher average success under tested shifts.
Juno connects robot predictions to motion and action learning
Juno links motion-aware robot predictions to action learning, with gains in tested simulation tasks and physical shifts.
Robot history helps most when video learning predicts forward
Robot tests find longer visual history helps after causal video pretraining, while bidirectional pretraining shows no net gain.
RealtimeWAM speeds robot action prediction while keeping success
RealtimeWAM coordinates computation in robot action models, lowering average simulation latency while keeping task success close to original inference.
Rules help frozen robot models handle new wording
A rule-based rewriter improved a frozen robot model on new instruction phrasings in simulated tasks.
Robot models learn from executed action alternatives to predict failures
CureWM repairs released robot world models with executed action alternatives, improving success–failure prediction in several tested settings.