A playground for automated RL env and physics simulations for Devin.
We've created playbooks (playbook/
) for enabling Devin to create new physics environments with MuJoCo.
- Run playbook
mujoco_wasm.devin.md
, take a snapshot to save sim environment - Modify and run
mujoco_create_env.devin.md
from snapshot for Devin to create a new simulation env (uses MuJoCo WebAssembly for live rendering.)
Render: node server.js
API:
-
/api/pendulum
POST for changing mass, length of rods, and initial parameters in pendulum.xml -
api/state
GET request for getting the current state of the Mujoco environment in a parsable JSON format (eg. current position x,y,z of the mass and timestep)
Devin is instructed to create an RL environment with 2 red balls and 1 blue ball, and automatically spins up a MuJoCo instance with live rendering/web access and makes a PR. (demo below, 4x speed)
demo.mp4
Built in a weekend with Devin technical preview (2nd place at Cognition hackathon)