A coding agent turns one trial into reusable skills: EmbodiedEvo scores 44.48 average on RoboDojo simulation
Synopsis
EmbodiedEvo lets a coding agent revise code through physical feedback on a single Playground scene, freeze the successful control procedures and failure conditions as reusable Skills and Lessons, and hand them to independent agents on new scenes. On RoboDojo-Sim, the Astra/xhigh configuration reaches 44.48 average score and 37.82% success rate, above all officially verified leaderboard results; in paired tests with task description alone, success rises from 8.50% to 30.17%, with gains concentrated on memory-dependent tasks while fine contact control remains difficult.
Interpretation
A robot-manipulation agent need not rediscover every operation in each new scene; code experience gathered in one scene can be frozen and carried to other layouts of the same task. Previously each coding-agent episode started with only the interface documentation and task description, and Astra reached just 8.50% average success on RoboDojo-Sim. In paired tests holding the executor and scenes fixed, EmbodiedEvo with experience raised success from 8.50% to 30.17% and average score from 14.77 to 38.64 when only the task description was supplied.
Experience is stored as two readable artifacts: Skills are executable control procedures, and Lessons record applicability conditions, failure modes and recovery guidance. Raw interaction history also preserves information, but the receiver must sift it from many programs and feedback records. With Astra/xhigh on matched test scenes, Skills/Lessons reached 38.64 score and 30.17% success versus 36.41 and 29.78% for raw history, while mean total cost per test episode fell from $12.203 to $5.732.
Acquiring experience and applying it are distinct capabilities, and experience produced by a stronger model lifts weaker executors. Harnesses that Sol and Luna evolved themselves brought little gain, with Sol's own harness at 2.50% success and Luna's at zero. Relative to no experience, Astra/xhigh's experience raised Sol/xhigh success by 8.86 percentage points and Luna/xhigh by 5.50, reaching 12.03% and 5.50% respectively.
Perspective
All evaluation runs in RoboDojo-Sim simulation; each task's harness is acquired on one Playground scene distinct from the test layouts, frozen as read-only files at test time, with GPT-6-Astra, GPT-5.6-sol and GPT-5.6-luna at the specified reasoning efforts. The full set uses 50 scenes per task with one seed, and paired analyses use a dev set of 5 scenes per task.
The paired runs illustrate mechanisms without isolating the causal effect of a single artifact such as a Skill or a Lesson; fine-contact tasks such as tube and key insertion still achieve no dev-set success, and success rates trail VPP2 and Simate-beta; the Skills/Lessons versus raw-history performance difference has confidence intervals spanning zero and is not a formal equivalence result; cost figures are estimates under a fixed rate schedule, not invoices.
