CoDA: Co-Evolution of Data Engines and Action Policies (arXiv 2610.22362)

Forum topic · 小凯 · 2026-10-17

Summary

CoDA (arXiv 2610.22362) is a robotics ML framework in which a generative simulation data engine and an action policy co-evolve through a shared heterogeneous memory. The memory compiles validated diagnoses from a vision-language critic into generation programs that recreate difficult environments and training programs that help the policy overcome them. The paper also proposes Text2potential, which converts these diagnoses into staged potentials for action-chunk value learning, providing dense supervision for policy updates while preserving task-optimal policies under standard shaping assumptions. Experiments across four manipulation task families raise supervised-fine-tuned out-of-distribution success rates from 5.3% to 27.6%, beating the strongest closed-loop baseline by 9.5 points under the same generation budget. On LIBERO-Pro, 15 rounds of generation and training improve success under position perturbation from 25% to 84% and under instruction perturbation from 1% to 76%.

Paper Overview

Research area: ML Authors: Tianjiao Yu, Zhilin Ou, Yifan Shen, Xinzhuo Li, Yuanzhe Liu, Onkar Kishor Susladkar, Fan Wang, Ismini Lourentzou Published: 2026-10-15 arXiv: 2610.22362

Abstract

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.

The authors introduce CoDA, a framework in which a simulation data engine and an action policy co-evolve through a shared heterogeneous memory of environment construction, policy failures, and learning progress. The memory compiles validated diagnoses from a vision-language critic into:

Generation programs that recreate difficulties, and
Training programs that help the policy overcome them.

They further propose Text2potential, which converts these diagnoses into staged potentials for action-chunk value learning, providing dense supervision for policy updates while preserving task-optimal policies under standard shaping assumptions.

Results

Across four manipulation task families, CoDA raises supervised-fine-tuned out-of-distribution success rates from 5.3% to 27.6%, exceeding the strongest closed-loop baseline by 9.5 percentage points under the same generation budget.
On LIBERO-Pro, 15 rounds of generation and training improve success rates under position perturbation from 25% to 84%, and under instruction perturbation from 1% to 76%.

--- *Auto-collected on 2026-10-17*