EMBODIED AIINSIGHTSThe systems beneath the motion
Clear analysis of models, data, sensing, control, and hardware without the hype layer.
AI-generated imageResearch brief
ForesightSafety-VLA separates successful robot behavior from safe behavior, revealing hazardous trajectories that conventional completion scores can count as wins.
AI-generated imageResearch brief
A humanoid policy trained across 150,000 simulated scenes reached a behavior-cloning ceiling. Reinforcement learning improved both simulated and zero-shot physical reach-and-pick performance.
AI-generated imageAnalysis
Humanoids cannot manipulate well if their legs merely preserve balance. New controllers coordinate locomotion, torso, hands and task-driven contacts through shared whole-body representations.
AI-generated imageResearch brief
A continual-learning VLA separates rapid skill adaptation from slower consolidation, then replays compact cached features instead of storing complete visual trajectories.
AI-generated imageAnalysis
Multi-stage robot work fails through accumulated errors, lost progress and weak recovery. New systems divide the problem by phase, tool and subgoal instead of merely predicting farther ahead.
AI-generated imageAnalysis
A robot that reaches the goal can still be brittle, unable to recover, or unsafe along the way. New benchmarks are beginning to measure those differences.
AI-generated imageResearch brief
SONIC scales a humanoid motion tracker across model size, motion data and compute, then connects the resulting controller to planners, teleoperation and vision-language-action commands.
AI-generated imageResearch brief
VistaVLA converts multi-view images into a semantic 3D Gaussian scene, then compresses that representation into 64 tokens for robot action prediction.
AI-generated imageAnalysis
Turning continuous motion into tokens lets language models control robots, but compression must balance precision, speed, semantics and transfer across different bodies.
AI-generated imageAnalysis
Touch can tell a robot about force, slip and hidden contact, but useful control requires separating slow semantic reasoning from rapid physical correction.
AI-generated imageResearch brief
A new vision-language-action model turns recent robot motion into compact temporal memories, improving multi-stage manipulation without running geometric reconstruction during deployment.
AI-generated imageAnalysis
Video world models can generate training data, preview possible futures and guide policies, but visual plausibility still does not guarantee correct physics or executable actions.