In simulation, a robot-training system raised average success on out-of-distribution task variations from 5.3% after supervised fine-tuning to 27.6% across four manipulation task families, according to a new preprint. It uses a policy’s failures to choose both future practice scenes and learning feedback.
The work, called CoDA, addresses a problem with generated training data: a scene can be physically valid but too easy, too hard or unrelated to a policy’s actual mistakes. Rather than generate variations blindly, the system looks for difficulties the policy can learn from and uses them to shape later rounds of practice.
Failure diagnoses steer both practice and training
After simulated attempts, a vision-language critic—a model that interprets images alongside text—uses rollout evidence to propose why the policy failed, such as an object being occluded during approach. Control scenes test the diagnosis before CoDA saves it in a shared memory.
CoDA then turns a validated diagnosis into two programs. One creates simulated scenes that reproduce the difficulty. The other breaks the task into subgoals and supplies intermediate learning signals, so the policy can receive feedback on progress before the whole task is complete. The system measures what improved and uses that record to choose later practice.
CoDA improved simulated success across two benchmarks
The authors report average out-of-distribution success of 27.6% for CoDA, compared with 5.3% after supervised fine-tuning and 18.1% for a closed-loop baseline guided by task success rates. The averages cover four manipulation task families and three random seeds; the 18.1% comparison uses the same environment-generation budget.
A separate LIBERO-Pro evaluation tested position and instruction perturbations after 15 rounds of generation and training. Reported success rose from 25% to 84% under position perturbations and from 1% to 76% under instruction perturbations. These are separate benchmark conditions, not one combined score.
The results are from an arXiv preprint and remain simulation results; they do not establish transfer to a physical robot. The authors have not reported an independent replication.
Sources
- CoDA: Co-Evolution of Data Engines and Action Policies, arXiv:2610.22362v1, submitted 15 October 2026.
- CoDA project page, PLAN Lab, with method and evaluation details.
Image credit: Hydraulic toy robot arm, by Scoid, CC BY-SA 4.0 via Wikimedia Commons. Illustrative; not used in CoDA.