CoDA: Co-Evolution of Data Engines and Action Policies
CoDA co-evolves robot simulation data and policies, lifting out-of-distribution success from 5.3% to 27.6%.
CoDA is a robot-learning framework in which a simulation data engine and an action policy co-evolve through a shared memory of environments, failures, and progress. A vision-language critic's diagnoses become generation and training programs, and Text2potential converts them into staged potentials for dense action-chunk supervision. 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. On LIBERO-Pro, 15 rounds raised success from 25% to 84% under position perturbations and from 1% to 76% under instruction perturbations.
- CoDA co-evolves a simulation data engine and an action policy via shared memory.
- Text2potential turns diagnoses into staged potentials for action-chunk value learning.
- Out-of-distribution success rose from 5.3% after fine-tuning to 27.6%.
- On LIBERO-Pro, 15 rounds lifted position-perturbation success from 25% to 84%.
- Products
- LIBERO-Pro
- AI models
- CoDA
From the original article
1 min read at arxiv.orgGenerative 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 · arxiv.org
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.