Remembering What Matters for Long-Horizon Robotic World Models
Autoregressive robot video world models based on discrete visual tokens remain unreliable over long rollout horizons.
ProofPaper ↗
Key points
- We investigate a possible memory-role conflict in Full Context decoding: persistent scene information, recent visual-action states, and stale generated history compete for attention despite their different predictive roles.
- This perspective motivates ReCAP (Receding Context with Anchor Preservation), a training-free context-scheduling method for long-horizon robot video world models.
- Across evaluated environments and horizons, ReCAP improves visual fidelity, action-conditioned consistency, and robot structure preservation relative to Full Context.
- Our results support inference-time context scheduling as a practical approach to improving long-horizon robot video prediction and computational efficiency without retraining.
Sources (1)
- [1]Remembering What Matters for Long-Horizon Robotic World ModelsarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 10, 05:01 AM
Autoregressive robot video world models based on discrete visual tokens remain unreliable over long rollout horizons.
We investigate a possible memory-role conflict in Full Context decoding: persistent scene information, recent visual-action states, and stale generated history compete for attention despite their different predictive roles.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 10, 2026BridgeWAM: Bridging Video and Action Experts via Latent Bridge Queries
- Oct 9, 2026PILOT: Guiding Long-Horizon World-Model Planning for Robot Manipulation with Visual Subgoals and Policy Priors
- Oct 9, 2026Dream to Remember: Continual Steering of Generalist Robot Policies with World Models
- Oct 9, 2026RSI-Anything: Unlocking Recursive Self-Improvement with World Models
- Oct 9, 2026Latent World Action Model for GUI Agents
- Oct 8, 2026Odyssey-3 is a new generative world model that you can try for free