CoDA: Co-Evolution of Data Engines and Action Policies
CoDA raised robot success on out-of-distribution tasks from 5.3% to 27.6%, beating the strongest closed-loop baseline by 9.5 percentage points.
Across four manipulation task families, that advantage used the same environment-generation budget as the baseline. The framework uses a shared memory of policy failures and learning progress to generate targeted simulation environments and training programs. Its Text2potential component turns diagnoses from a vision-language critic into dense supervision for policy updates. On LIBERO-Pro, 15 generation-and-training rounds raised success from 25% to 84% under position perturbations and from 1% to 76% under instruction perturbations.
HF Daily Papers' note
Across four manipulation task families, that advantage used the same environment-generation budget as the baseline. The framework uses a shared memory of policy failures and learning progress to generate targeted simulation environments and training programs. Its Text2potential component turns diagnoses from a vision-language critic into dense supervision for policy updates. On LIBERO-Pro, 15 generation-and-training rounds raised success from 25% to 84% under position perturbations and from 1% to 76% under instruction perturbations.
HF Daily Papers' note
score 5