robotics
Nineteen Seconds of History Finally Changed the Robot's Next Move
Long-WAM gained from longer context only when its video foundation had learned causal prediction rather than bidirectional reconstruction.
Summary
Long-WAM gained from longer context only when its video foundation had learned causal prediction rather than bidirectional reconstruction.
Long-WAM asks whether a real-time robot can use more visual history without delaying control. Its video foundation is pretrained autoregressively on robot and egocentric footage, then adapted while preserving the history-to-future objective. On RoboCasa GR-1, extending context from zero to 19.2 seconds raised success from 63.3% to 78.7%; a bidirectionally pretrained initialization gained nothing overall. Streaming encoding and asynchronous execution kept an action chunk, including predicted future-video latents, to 107.4 milliseconds on an RTX 5090. The authors also report 95% success on dynamic cup stacking with a Unitree G1 where two compared policies failed all twenty trials. These are benchmark and controlled hardware results, but they isolate a useful distinction: storing history is not the same as training a controller to use it causally.
Why it matters
Long-WAM gained from longer context only when its video foundation had learned causal prediction rather than bidirectional reconstruction.
Limits and context
- These are benchmark and controlled hardware results, but they isolate a useful distinction: storing history is not the same as training a controller to use it causally.
Key claims
Long-WAM gained from longer context only when its video foundation had learned causal prediction rather than bidirectional reconstruction.
Qualification: These are benchmark and controlled hardware results, but they isolate a useful distinction: storing history is not the same as training a controller to use it causally.
Evidence: source-2026-10-08-008
Sources
- arXiv preprint 2610.10528arXiv · primary research
Corrections
No corrections have been recorded for this story.