Robot policies preserve pretrained knowledge and delegate recovery to lightweight control layers

Two independent methods identify representation loss during behavior-cloning fine-tuning as a generalization failure. Semantic Anchoring aligns a shared action channel to a frozen text manifold and finds that alignment tracks out-of-distribution success with Spearman ρ=0.964. Anchor-Align instead distills frozen vision-language hidden states and adds motion-direction prediction on robot observations. It reaches 71.9% on LIBERO-PRO, versus 61.0% for its VLA-Adapter baseline. Together, the studies support preserving pretrained concepts while leaving room for execution-specific features.

GigaWorld-Policy-0.5 learns from future visual dynamics but skips video generation at deployment through separate visual and action experts; its action-only path runs at about 85 ms on an RTX 4090. A complementary execution manager keeps the underlying manipulation policy frozen and chooses among execute, retry, repair, and reset modes. Under injected disturbances, it improves LIBERO suite success by 25.7 to 39.2 percentage points. Both designs isolate costly prediction or recovery logic from the core action generator rather than expanding every inference step.

PhysClaw-0 stores natural-language corrections as reusable rules, so recurring collection failures do not require repeated operator intervention. It collected 50 valid demonstrations with 4.8 minutes of human work, versus 30 minutes for teleoperation, while matching the downstream policy’s 80% success rate. The Industrial Dexterity Benchmark shows the parallel value of richer sensing: multimodal imitation learning reached 78% on a cable grasp-and-insert task, compared with 36% for single-camera RGB. These are narrow evaluations, but they show how persistent corrections and task-specific sensing can reduce dependence on a monolithic policy.