robot-agent-exploration/skills-lib

---

RAX is the general harness: a coding agent (Claude Code or Codex CLI) drives a robot through perception and control tools, with no demonstrations, and compiles what worked into code. This repository keeps what those agents learned on real robots.

Each skill is an executable procedure the agent wrote while it explored a task, stored with its preconditions, checks, failure modes and the evidence of its runs (run records and camera frames). A later session, on a new task or in a new scene, reads the library, runs or composes the skills that match, and merges back what it improves.

 explore a task ──► write a skill ──► merge into the library
        ▲                                     │
        └──── next session reads, reuses ◄────┘
              and composes the skills

🌿 Libraries

main holds the protocol. Each robot embodiment has its own branch, with one folder per skill.

Each library’s harness lives in the RAX repository, which also documents the robot’s tools: real/rainbow and real/franka.

🚀 Using a skill

git clone -b rby1 https://github.com/robot-agent-exploration/skills-lib.git skill-library

1. Read INDEX.md and open only the skills that match your task.
2. Check each skill’s preconditions against your scene.
3. Run its entry through the robot’s harness context (ctx).

Inside RAX this happens automatically: the harness clones the robot’s library into each session’s workspace and gives the agent the continual-learning skill to read, run and grow it. RAX’s real-robot adaptation guide sets this up for every new robot; real/rainbow is the reference.

📁 Layout of a library

README.md                    the robot and where its skills came from
INDEX.md                     one row per skill: status, what it does, when to use it
skills//
  SKILL.md                   when to use, preconditions, procedure, code, checks,
                             failure modes, evidence
  code/run.py                the entry: main(ctx, ...) -> {ok, measured, why}
  code//            its modules, a package named after the skill
  evidence/runs.jsonl        one record per run of the code: outcome, unattended, conditions, checks
  evidence/images/           frames of those runs

A skill’s status follows its evidence:

🤝 Contributing

CONTRIBUTING.md covers what to contribute, how to keep skills modular so new tasks can compose them, the format and evidence rules, and how to commit and merge into an embodiment branch. Check a contribution with scripts/check_skills.py before opening a pull request; the embodiment branches run it on every push and pull request. To bring a robot that has no branch yet, follow NEW_ROBOT.md.

📝 Citation

These skills come from sessions of RAX: Scaling Online Exploration for Robot Agents. If you use them, please cite:

@misc{huang2026scaling,
  title        = {Scaling Online Exploration for Robot Agents},
  author       = {Huang, Runhan and Zhang, Xiangcheng and Chen, Haoyu
                  and Li, Yitong and Wang, Yen-Jen and Lyu, Bohan
                  and Ye, Haotian and Hong, Chuye and Wang, Yixuan
                  and Su, Chloe and Xiong, Haoyu and Wu, Yangzhen
                  and Xie, Enze and Manderson, Travis and Ermon, Stefano
                  and Abbeel, Pieter and Torralba, Antonio and Song, Dawn
                  and Ye, Weirui and Du, Yilun},
  year         = {2026},
  howpublished = {Preprint},
  url          = {https://robot-agent-exploration.com/}
}