Robot agents and VLA policies stop training apart: RoboCoEvo turns failure programs into policy data, lifting success across three benchmarks
Synopsis
RoboCoEvo closes the loop between a robot agent and a vision-language-action (VLA) policy: the agent groups recurring failures into failure modes, writes one reusable detector and recovery program per mode, and uses those programs on failure states and perturbed copies to generate corrective data that DAgger-style post-training writes into the policy. The updated policy's shifted failure distribution then guides the next round of program revision. Across two co-evolution rounds on BEHAVIOR-1K, RoboTwin and RoboCasa, success rises from 0.23 to 0.34, from 0.79 to 0.93 and from 0.79 to 0.91.
Interpretation
In a robot system, the agent that diagnoses and recovers failures and the VLA policy that executes low-level actions can now push each other: recovery programs written by the agent become policy training data, and the failures left after policy improvement decide what the agent revises next. In earlier combined systems, agent improvement and policy improvement stayed largely separate, so gains on one side did not necessarily benefit the other; a policy that does not absorb corrective experience keeps depending on the agent, and an agent not updated with the policy keeps targeting weaknesses that no longer limit performance. On the three simulated benchmarks BEHAVIOR-1K, RoboTwin and RoboCasa, average success after two co-evolution rounds rises from 0.23 to 0.34, from 0.79 to 0.93 and from 0.79 to 0.91, and the mean BEHAVIOR-1K Q-score rises from 0.40 to 0.56, with gains in each round.
Failures are grouped into failure modes, defined as stalling at the same subtask with the same pattern in the scene-state sequence around the failure, and each mode gets a single detector and a single recovery program written as a finite-state machine that composes navigation and manipulation skills. Diagnosing and recovering each failed rollout separately makes cost grow with the number of failures and can yield inconsistent recovery strategies for similar failures, producing high-entropy demonstration data. On the BEHAVIOR-1K picking up trash task, per-failure recovery lowers Q-score from the 0.633 baseline to 0.267, while sharing one program per failure mode raises it to 0.758 and cuts agent token usage to about one sixteenth.
Detected failure states are perturbed in the simulator to create more failure scenes of the same mode, the recovery programs run on those states, and only successful recoveries are kept as corrective experience for training. Each rollout yields only a few instances of a given failure mode, so collecting failure states by adding rollouts raises simulation cost. On picking up trash, full RoboCoEvo obtains failure states from 50 rollouts plus perturbation in 36 GPU hours, versus 147 GPU hours for the failure-mode variant without augmentation, and Q-score rises from 0.758 to 0.84.
Perspective
The demonstrated scope is simulated evaluation on the first six BEHAVIOR-1K tasks, RoboTwin 2.0 and ten RoboCasa tasks, starting from RLC, a VLA trained on clean and randomized demonstrations, and XR-1 respectively, with GPT-6 Astra as the agent; the updated policy is deployed without the agent.
Learning currently uses only successful recovery demonstrations, while the VLA's own successful and failed rollouts are not used for training; the authors list combining RoboCoEvo with reinforcement learning to exploit those rollouts as future work. Gains are limited on some tasks: on picking up trash the Q-score rises from 0.63 to 0.84 after the first round and dips slightly to 0.82 in the second, which the authors read as possible saturation.
