CoDA co-evolves simulation and robot policies via shared failure memory, lifting OOD success from 5.3% to 27.6%

PLAN-Lab · hf · 2026-10-16

Researchers from UIUC, CUHK-Shenzhen, and Amazon introduce CoDA, a framework where a generative simulation data engine and a robot action policy co-evolve, tackling the gap between physically valid environments and environments that are actually useful for training.

Shared memory: the engine and policy share a heterogeneous memory of environment construction, policy failures, and progress; a vision-language critic compiles diagnoses into generation programs that recreate difficulties and training programs that help the policy overcome them.
Text2potential: converts diagnoses into staged potentials for action-chunk value learning, providing dense supervision while preserving the optimal policy under standard shaping assumptions.
Results: across four manipulation task families, out-of-distribution success improves from 5.3% after supervised fine-tuning to 27.6%, beating the strongest closed-loop baseline by 9.5 points with the same generation budget; on LIBERO-Pro, 15 rounds raise success from 25% to 84% under position perturbations and from 1% to 76% under instruction perturbations.

Original post →

CoDA co-evolves simulation and robot policies via shared failure memory, lifting OOD success from 5.3% to 27.6%

More from Embodied