LeRobot Simulation Environment

AUTHOR: Kevin Thomas

CREATION DATE: January 11, 2026
UPDATE DATE: February 28, 2026

Train a robot brain from zero to 100% success in 60 seconds flat — no hardware, pure NumPy, pure magic. Built on HuggingFace LeRobot standards.

Install

python3 -m venv .venv && source .venv/bin/activate
pip install ".[dev]"

Train

python run_sim.py --task pusht --mode train

Collects demo episodes with a scripted expert, trains an MLP via behaviour-cloning (Adam optimiser), then evaluates. Saves best_policy.npz (lowest loss) and last_policy.npz to --output-dir.

Available tasks: pusht, pick_place, reach

Flag	Default	Description
--demo-episodes	20	Expert demo episodes to collect
--train-steps	3000	Behaviour-cloning gradient steps
--eval-episodes	5	Evaluation rollouts after training
--hidden-dim	256	MLP hidden layer width
--output-dir	./sim_output	Where to save models and data
--seed	42	Random seed

Run Inference

python run_sim.py --task pusht --mode eval

Loads best_policy.npz from --output-dir and runs evaluation episodes. Requires --mode train first.

Teleoperate

python run_sim.py --task reach --mode teleop

Drive the arm with WASD / arrow keys and record demonstrations.

Visualize

python run_sim.py --task pick_place --mode visualize

PyGame live rendering of a random policy.

Environments

Task	DOF	Action	Observation
pusht	2-D	(dx, dy)	6-D state (agent + block + target) + optional RGB
pick_place	3-D	6 joint vels + 1 gripper	13-D state + optional 384x384 RGB
reach	3-D	6 joint vels	9-D state + optional 384x384 RGB

Project Structure

run_sim.py                  CLI entry point
lerobot_sim/
  envs/                     Gymnasium environments + factory + configs
  policies/                 BasePolicy, RandomPolicy, Expert policies, MLPPolicy, Trainer
  datasets/                 Episode recording + HF Hub loader
  robots/                   6-DOF simulated arm (SO-100 style)
  teleop/                   Keyboard and gamepad input
  visualization/            PyGame renderer + HUD + replay
  utils/                    Constants + helpers

LeRobot Compatibility

Observations use observation.state / observation.image keys. The factory returns {suite: {task_id: VectorEnv}} matching lerobot.envs.factory.make_env. Datasets are NumPy archives convertible to Parquet + MP4 for Hub upload.

EnvHub

This project includes an env.py at the root so it can be loaded directly via LeRobot EnvHub:

from lerobot.envs.factory import make_env
envs = make_env("your-user/lerobot-sim", trust_remote_code=True)

Resources

• LeRobot Robot Learning Tutorial — free hands-on course covering teleoperation, dataset recording, policy training, and inference
• LeRobot Documentation — full API docs, installation, and hardware guides
• LeRobot GitHub — upstream source code

Arduino UNO Q Deployment

The lerobot-policy-deployment-app/ folder contains a complete Arduino App Lab application that deploys the trained MLP policy to an Arduino UNO Q. Enter state observations through a web interface, and the board drives servos in real time using the NumPy forward pass while displaying episode status on the 8 x 13 LED matrix. See lerobot-policy-deployment-app/README.md for full details.

Author

Kevin Thomas

• Creation Date: January 11, 2026
• Last Updated: February 28, 2026

License

Apache 2.0