Robotics · 2026-10-16
CoDA: Co-Evolution of Data Engines and Action Policies
Tianjiao Yu, Zhilin Ou, Yifan Shen, Xinzhuo Li, Yuanzhe Liu, Onkar Kishor Susladkar, Fan Wang, Ismini Lourentzou
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
CoDA co-evolves a simulation data engine and an action policy via a shared heterogeneous memory of scene-construction traces, rollout-grounded failure diagnoses, and learning progress. A vision-language critic produces validated diagnoses that are compiled into (i) generation programs to recreate difficulties and (ii) Text2potential-based training programs that provide dense potentials for action-chunk value learning. Across four manipulation task families, CoDA improves OOD success to 27.6% from 5.3% after supervised fine-tuning.
Why it matters
The authors argue that generating physically valid environments is not enough, and that a generator needs feedback about what difficulties the policy can still learn from. CoDA uses validated, factor-/phase-local failure diagnoses plus measured learning progress to adapt what environments to generate and how to train, aiming to improve out-of-distribution success without relying on human reward engineering.
Method
- Co-evolution via shared heterogeneous memory: episodic construction/rollout traces, semantic difficulty modes, and procedural generation/training programs indexed by factor and trajectory phase.
- Generation programs reproduce targeted failure conditions; training programs are compiled from diagnoses using Text2potential to form staged potentials for action-chunk value learning.
- Memory-guided scene generation uses hardness (success rate) and learning progress; control experiments compare targeted difficulty modes vs control factor variants before compiling.
Limitation
CoDA relies on foundation models for scene generation, failure diagnosis, and potential compilation. Errors in these models’ understanding can affect scenes and training signals; additional simulation and model inference are required, constraining the number of rounds and task families under a fixed budget. Evaluation focuses on simulated manipulation; transfer to physical robots remains to be established.
Abstract (from arXiv)
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. We 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. We further propose Text2potential to convert these diagnoses into staged potentials for action-chunk value learning, providing dense supervision for policy updates while preserving the task's optimal policy under standard shaping assumptions. Measured learning progress guides further generation and program refinement. Across four manipulation task families, CoDA improves out-of-distribution success from 5.3% after supervised fine-tuning to 27.6%, exceeding the strongest closed-loop baseline by 9.5 percentage points with the same generation budget. On LIBERO-Pro, 15 rounds of generation and training raise success from 25% to 84% under position perturbations and from 1% to 76% under instruction perturbations.
Related papers
- SkillWeaver: Agentic Exploration over Neural Interaction Skills for Scalable Robot Data Generation
- RAP-WAM: Learning World-Action Models through Predictive Rehearsal and Action Perturbations
- Active Generative Simulation: Meta-Learning What to Simulate for Sim-and-Real Policy Co-Training
- Memory Without Curation: Learning Multi-Stream Latent Memory for Robot Policies
- Don't Throw Away the Tail: Action Upcycling for Policy Acceleration
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-15 · Indexed by Pub-AI: 2026-10-16 UTC
CoDA: Co-Evolution of Data Engines and Action Policies
Preprint posted: 2026-10-15 · Original record ↗
- arxiv
- 2610.22362
Explicit source versions (1)
- v1 · 2026-10-15
- Preprint available
Source date: 2026-10-15 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
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. We 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. We further propose Text2potential to convert these diagnoses into staged potentials for action-chunk value learning, providing dense supervision for policy updates while preserving the task's optimal policy under standard shaping assumptions. Measured learning progress guides further generation and program refinement. Across four manipulation task families, CoDA improves out-of-distribution success from 5.3% after supervised fine-tuning to 27.6%, exceeding the strongest closed-loop baseline by 9.5 percentage points with the same generation budget. On LIBERO-Pro, 15 rounds of generation and training raise success from 25% to 84% under position perturbations and from 1% to 76% under instruction perturbations.