robotics
The Robot Keeps the Controller and Relearns the View
A frozen action decoder transfers privileged reinforcement-learning behavior into visual manipulation with fewer demonstrations.
Summary
A frozen action decoder transfers privileged reinforcement-learning behavior into visual manipulation with fewer demonstrations.
Latent RL Prior first distills simulated reinforcement-learning rollouts into a flow-matching action decoder conditioned on proprioception and latent task context. Visual adaptation freezes that decoder and trains only an image-to-context interface, with no privileged labels or latent targets. Across three simulated tasks, the method beat full-parameter fine-tuning in the low-data regime and surpassed full-data baselines on door lifting and bimanual assembly using ten and five times fewer demonstrations respectively. The evidence is simulation-based, but it isolates a practical route for reusing control knowledge.
Why it matters
A frozen action decoder transfers privileged reinforcement-learning behavior into visual manipulation with fewer demonstrations.
Limits and context
- Visual adaptation freezes that decoder and trains only an image-to-context interface, with no privileged labels or latent targets.
Key claims
A frozen action decoder transfers privileged reinforcement-learning behavior into visual manipulation with fewer demonstrations.
Qualification: Visual adaptation freezes that decoder and trains only an image-to-context interface, with no privileged labels or latent targets.
Evidence: source-2026-11-08-007
Sources
- arXiv preprint 2611.08458, original v1arXiv · primary research
Corrections
No corrections have been recorded for this story.