CoDA: Co-Evolution of Data Engines and Action Policies
CoDA co-evolves simulation data engines and robot policies, lifting out-of-distribution manipulation success.
CoDA couples a generative simulation data engine with an action policy through shared memory of environment construction, failures, and learning progress. A vision-language critic's diagnoses become generation programs and Text2potential staged rewards for action-chunk value learning. Across four manipulation families, out-of-distribution success rose from 5.3% after supervised fine-tuning to 27.6%, beating the strongest closed-loop baseline by 9.5 points at the same generation budget. On LIBERO-Pro, 15 rounds raised success from 25% to 84% under position perturbations and from 1% to 76% under instruction perturbations.
- Shared memory links environment generation to policy failures
- Text2potential turns diagnoses into staged value-learning potentials
- Out-of-distribution success rose from 5.3% to 27.6%
- LIBERO-Pro success reached 84% and 76% after 15 rounds
From the original article
1 min read at huggingface.co
Generative simulation can automate robot environment construction, but physically valid environments are not necessarily useful for training. The challenge is to identify environments that address a policy's failures and supervision that helps it overcome them.
By Tianjiao Yu, Zhilin Ou, Yifan Shen · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.