A reproducible environment for policy self-improvement towards automated biological research.
The environment is four components:
| component | where |
|---|---|
| Skill | AGENTS.md — operating rules and the starting workflow for the agent |
| Code exemplar | branch wetrobo+petri+cap+door — the repository after a coding agent evolved it in this environment through three trials (Petri dish lid, culture-media bottle cap, incubator door) |
| Arm | one AgileX Piper arm with a string-driven Dynamixel gripper, a fixed head RGB-D camera (iPhone via Record3D), and a wrist RGB camera |
| Incubator | a laboratory incubator in front of the arm |
| Demonstrations | demo/ — fifteen teleoperated demonstrations per task |
The task is not part of the environment; the operator gives it at run time.
demo/ holds fifteen teleoperated demonstrations per task, recorded on this
equipment with a Meta Quest interface:
demo/petri_lid/ lifting the Petri dish lid
demo/bottle_cap/ lifting the bottle cap
demo/incubator_door/ opening the incubator door
Each episode is one *.hdf5 trajectory plus a head-camera preview
*_head.mp4. See demo/README.md for the datasets and a replay command.
main— wetrobo, the base control stack the agent starts from: one-demo replay, live bias, safety layer, camera streaming, and the arm RPC server. This is the tree the trials began with.wetrobo+petri,wetrobo+petri+cap,wetrobo+petri+cap+door— byte-exact snapshots of the working repository at the end of each trial, in the order the trials ran (Petri dish lid, then bottle cap, then incubator door). They form a chain, sogit diffbetween adjacent branches shows what one trial added.wetrobo+petri+cap+dooris the code exemplar: the final form of the three task programs at the paths where they run.
Declared in pyproject.toml. In addition, the code imports:
piperlib— AgileX Piper CAN driver used byrobot/arm/arm.pyrecord3d— iPhone RGB-D streaming used byrollout/camera.pyscipy,opencv-python,h5py
robot/camera_map.json, src/configs/safety.json, and the MJCF under
robot/cone-e-description/ describe one installation; measure your own before
commanding motion.
Additional demonstration data (3,000+ trajectories) collected during our study is available on Hugging Face:
- Playingyoyo/wetrobo-related (Note: These demonstrations were not directly used in WetRobo).