Long-WAM Scales Robot World-Action Context
First coverage.
Long-WAM raises robot success by extending causal world-action context after autoregressive video pretraining.
Long-WAM scales the context of causal world-action models for real-time robot control. On RoboCasa GR-1, extending visual context from 0.0 to 19.2 seconds raises success from 63.3% to 78.7% when the video foundation is pretrained autoregressively, while a bidirectional initialization shows no net gain. Both reports say it leads compared methods on LIBERO-Long, RoboTwin 2.0, and DOMINO, and that each action chunk, including future-video latent prediction, takes 107.4 ms on an RTX 5090. Dynamic cup stacking reaches 95% success, whereas Pi0.5 and Fast-WAM succeed in none of 20 trials. The sources agree on Unitree G1 and YAM deployment; Hugging Face also lists DGX Spark and Jetson AGX Thor, which the arXiv report does not mention.
- Autoregressive video pretraining makes longer visual history useful for control; a bidirectional initialization shows no net gain.
- On RoboCasa GR-1, extending context from 0.0 to 19.2 seconds raises success from 63.3% to 78.7%.
- Long-WAM leads compared methods on LIBERO-Long, RoboTwin 2.0, and DOMINO.
- Dynamic cup stacking reaches 95% success; Pi0.5 and Fast-WAM succeed in none of 20 trials.
- Each action chunk, including future-video latent prediction, takes 107.4 ms on an RTX 5090.
- Deployment is reported on Unitree G1 and YAM; the Hugging Face report also lists DGX Spark and Jetson AGX Thor.
Coverage
oldest first · one row per article · UTC- · 24d ago
Hugging Face daily papersFirst report
Long-WAM: Scaling the Context of World-Action Models
Long-WAM raises robot task success by scaling causal world-action context with autoregressive video pretraining.
- · 23d ago
arXiv cs.AI / cs.LG / cs.CL
Long-WAM: Scaling the Context of World-Action Models
Long-WAM scales robot world-action context, lifting RoboCasa success from 63.3% to 78.7%.