MotionVLA accepted at CoRL 2026: remember motion, not frames
yshan2u · x · 2026-09-08
MotionVLA, from Xinggang Wang's team, is accepted at CoRL 2026 (arXiv:2606.08288).
Core claim: VLA models should remember the motion connecting past frames, not the frames themselves. Existing approaches condition on history, depth, or 4D features to resolve long-horizon ambiguity, but motion-inconsistent evidence introduces geometric drift, fragmented temporal cues, and unstable action generation.
Method:
- Converts a short past-only video window into compact, time-continuous trajectory-field tokens;
- Current visual tokens query this history to retrieve task-relevant motion information;
- The retrieved motion is recoupled into the VLA stream under trajectory-grounded supervision.
Across simulation benchmarks and preliminary real-robot rollouts, MotionVLA improves long-horizon manipulation with smoother, more direct executions — effective VLA memory is about usable motion-consistent evidence, not more 4D context.
More from Embodied
- RealSense D585 Pro AI-Powered 3D Stereo Depth Camera Unboxed for Physical AI — chrismatthieu · 2026-10-26
- Physical world agent swarm demo shared on X — pduan · 2026-10-26
- Looking for a wearable mic with best battery life that feeds recordings to personal agents via MCP — altryne · 2026-10-26
- 'Lamp Swarm': a desk of lamp-shaped robots shows what physical AI work looks like — dee_hw · 2026-10-26
- Tencent's HiSTEVE 1.0: hierarchical VLA agent plays Minecraft with native keyboard and mouse, hits 85% SOTA — _akhaliq · 2026-10-26
- SJTU's VaM lets robots learn online from mistakes, 95%+ success over 4-hour autonomous runs — chris_j_paxton · 2026-10-26