CoDA co-evolves sim engine and policy, OOD rises to 27.6%
TL;DR
- CoDA lifts out-of-distribution success on four manipulation task families from 5.3% after supervised fine-tuning to 27.6%.
- On LIBERO-Pro, 15 rounds of generation and training take instruction-perturbation success from 1% to 76% and position perturbations from 25% to 84%.
- A vision-language critic diagnoses policy failures, then writes both new environments and training programs into a shared memory.
Out-of-distribution success on four robot manipulation task families rose from 5.3% after supervised fine-tuning to 27.6% under a training loop called CoDA, according to a paper on arXiv by Tianjiao Yu, Ismini Lourentzou and six co-authors.
The setup pairs a generative simulator with the policy that will eventually run in it. "A simulation data engine and an action policy co-evolve through a shared heterogeneous memory of environment construction, policy failures, and learning progress," the abstract states. A vision-language critic validates diagnoses of what the policy is failing at; those diagnoses then get compiled into both new environments that recreate the difficulty and training programs that help the policy overcome it. A method the authors call Text2potential turns the same diagnoses into staged potentials for action-chunk value learning.
The largest reported gains are on LIBERO-Pro. After 15 rounds of generation and training, success climbed from 25% to 84% under position perturbations and from 1% to 76% under instruction perturbations. Against the strongest closed-loop baseline at the same generation budget, CoDA ran 9.5 percentage points higher.
The abstract names no specific robot platform and reports no real-hardware trial.
Originally reported by paper
Read the original article →Original headline: UIUC CoDA Lifts Robot OOD Manipulation 5× via Co-Evolving Sim Engine and Policy