EmbodiedEvo: Self-Evolving Coding Agent For Physical Intelligence

Observe, revise code, and act. Learn Skills and Lessons from physical feedback.

EmbodiedEvo connects a coding agent's workspace to a persistent RoboDojo simulation session. The agent inspects observations, edits Python control code, executes it, and uses the resulting physical feedback to guide its next revision. Its evolving knowledge lives in reusable code and written lessons in the workspace.

1. Observe, revise, execute. Inspect camera observations and robot state, update control code, and execute the revised segment in the same Python and simulation session. New observations close the interaction loop.
2. Build Skills and Lessons in Playground. Turn interaction experience into reusable helpers in skills/ and practical guidance in lessons/, such as grasping from above, verifying a grasp, or carrying an object in small steps. Playground permits same-scene resets while the agent refines this knowledge.
3. Freeze and reuse in Test. Freeze only skills/ and lessons/ before selecting Test scenes. Each scene starts with a fresh agent, workspace, and simulator. The transferred knowledge is read-only; the agent can observe and revise code within the scene, but cannot reset or learn across Test scenes.

Use a Linux host and Python 3.11 or newer for the orchestration environment:

python3 -m venv .venv
.venv/bin/python -m pip install -r requirements-host.txt -e '.[dev]'
.venv/bin/python -m unittest discover -s tests -v

Build the thin coding-agent image directly from the included Dockerfile:

docker build -t embodiedevo-agent:latest docker/agent

The Dockerfile installs pinned Codex CLI 0.159.2 and basic Python/image tools; no host Codex installation or API key is needed for the build. Its context contains no credentials, task files or simulator assets. Build the image on each Docker host that will run agents; the simulator is provisioned separately. See the complete image build procedure.

Run commands from this repository root. The editable installation keeps task templates and recipes available at their source paths. The project is currently distributed as a source checkout, not a standalone wheel with simulator assets.

Real episodes require a separately provisioned RoboDojo/Isaac Sim environment, assets, an available NVIDIA GPU, Docker, and configured model credentials. Follow installation, including Codex API configuration, before starting an experiment.

After setup, choose a native task with --env-task. The model and reasoning effort are explicit in tasks/RoboDojo/task.toml; configure them before a new run.

RUN_DIR="runs/$(date +%Y%m%d%H%M)_stack_bowls"
.venv/bin/embodiedevo explore --task RoboDojo --env-task stack_bowls \
  --run-dir "$RUN_DIR" --gpu 0
.venv/bin/embodiedevo test --task RoboDojo --env-task stack_bowls \
  --run-dir "$RUN_DIR" --gpu 0

embodiedevo run performs both stages. Test-only baseline:

.venv/bin/embodiedevo test --task tasks_baseline/RoboDojo --env-task stack_bowls \
  --run-dir "runs/$(date +%Y%m%d%H%M)_stack_bowls_baseline" --gpu 0

The baseline starts with its own instruction and no transferred Skills/Lessons. Public task recipes are enabled by default for both configurations, matching the source snapshot. Set EMBODIEDEVO_USE_RECIPE=0 before preparing a new run to omit that supplied knowledge. The choice and recipe source are recorded per run. See the protocol and running and evaluation.

Cluster campaigns, private deployment profiles, historical recovery tools, dashboards and result archives are outside this source distribution. Source provenance describes the export and its validation limits.

A project-wide license has not yet been selected. This is a local release preparation, not yet a licensed public release. The included RoboProbe task recipes retain their Apache-2.0 license and source attribution; see third-party notices.

If you use EmbodiedEvo in your research, please cite our paper.

@misc{li2026embodiedevoselfimprovingcodingagents,
  title         = {EmbodiedEvo: Self-improving Coding Agents for Physical Intelligence},
  author        = {Wenyi Li and Yizhe Chi and Youjie Zheng and Chunyu Liu and
                   Xiaoqiu Wang and Mingju Gao and Zheng Jiang and Deyao Hong and
                   Calvin Xiao and Kaisen Yang and Qinhuai Na},
  year          = {2026},
  eprint        = {2610.18867},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2610.18867}
}