Think Like a World Model, Act Like a VLA: Distilling World-Model Representations into Compact Robot Policies
Think Like a World Model, Act Like a VLA: Distilling World-Model Representations into Compact Robot Policies
This research proposes a method to distill world-model representations into compact robot policies by aligning student policies with a frozen world model during training. The approach achieves performance improvements on tasks like LIBERO and RoboCasa-GR1, with a 0.8B student reaching 97.9% on LIBERO and improving from 48.2% to 50.5% on RoboCasa-GR1. The method is also effective on real hardware, demonstrating broad representational prior rather than fragile network alignment.
This paper introduces a method to distill world-model representations into compact robot policies, improving performance on tasks like LIBERO and RoboCasa-GR1 without requiring a teacher during training. The approach leverages a frozen world model to align student policies, achieving significant gains in efficiency and task success.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
ith that cache. No teacher is loaded during training, the projector is discarded after it, and the deployed policy is identical to the undistilled baseline, running in 32ms and 1.86GB on a consumer RTX5090, so every gain is attributable to the representation rather than to added capacity or test-time compute. A 0.8B student reaches 97.9% on LIBERO, improves from 48.2% to 50.5% on RoboCasa-GR1 humanoid manipulation, a
Open source S4
Source：arXiv Robotics — research abstracts · arxiv.org