A robot-coding AI agent carried control procedures learned in one simulated scene into unseen layouts in a manipulation study. On matched tasks that gave the system only a task description, it succeeded in 30.17% of episodes, compared with 8.50% for the same model without saved experience.
The system, EmbodiedEvo, addresses a problem that appears when every new attempt starts with a fresh agent: useful corrections from one scene may be lost before the next. A robot may need more than a correct target. It also needs a reliable way to grasp an object, check that the grip held and know when it has gathered enough information to act.
In a simulated “Playground,” the agent writes and runs Python control programs, then revises them after observing the robot’s state and the result of each action. EmbodiedEvo stores two kinds of experience. “Skills” are reusable procedures, such as a grasp-and-lift routine. “Lessons” record when a procedure applies, what failed and what the agent should check or try next.
After exploration, the system freezes those files. A fresh agent then receives them as read-only guidance while handling a different layout of the same task. The model weights, task goals and robot interface stay fixed; the test agent can still adapt its code to the new scene. This is experience transfer through saved programs and notes, not fine-tuning the language model.
The 30.17% result comes from a matched development-set comparison: five scenes per task and 210 episodes per condition, with GPT-6 Astra using xhigh reasoning. The same scenes and task-description-only input were used with and without Skills and Lessons. In a separate full-benchmark run, the authors tested 2,100 episodes across 42 RoboDojo-Sim tasks, using task descriptions plus the benchmark wiki. They reported a 37.82% success rate and an average score of 44.48 on a 0–100 scale, calculated by giving each of the five task categories equal weight. The score gives credit for partial task completion under task-specific rules; success rate counts episodes meeting each task’s completion criterion. The full-set evaluation used one seed, so the paper does not show how results vary across repeated runs. The two headline results use different test scales and input settings.
Performance varied by task. In the full-set results, success was 13.25% on Precision tasks, which require accurate contact, and 23.50% on Long-Horizon tasks, where success depends on completing a sequence of actions. The paper describes grasps that reached a commanded position without securely holding an object, and multistep tasks that remained unfinished. Reusable code can help an agent remember what to do; it does not ensure that a delicate physical action succeeds.
The paper also compared its compact Skills and Lessons with giving the agent its raw interaction history. On the matched development scenes, success was similar: 30.17% versus 29.78%. The authors estimate that structured experience cost $5.73 per test episode, compared with $12.20 for raw history, using fixed API prices. Those are modeled estimates, not invoices.
The paper reports EmbodiedEvo results only in simulation and includes no physical-robot trials.
Figure 1 from Wenyi Li and colleagues’ EmbodiedEvo paper, licensed under CC BY 4.0.