robot-template

SmolVLA driving the LIBERO benchmark over HUD's robot capability. environment/env.py declares the env, environment/lerobot_sim.py is the module-level sim factory (task is a build arg), and run.py is the agent and runner.

Local

# once — EGL/Mesa + cmake (LIBERO pulls robomimic, which builds an EGL probe)
sudo apt-get install -y cmake libgl1 libosmesa6-dev libglew-dev libegl1 libglib2.0-0 patchelf

uv sync
MUJOCO_GL=egl uv run python run.py

Edit SUITE / N_TASKS in run.py to pick the suite and task count. With HUD_API_KEY set, rollouts stream to the trace viewer on hud.ai.

Docker (no local LIBERO)

Build the sim image once, then point the runner at it — the policy still runs on the host:

docker build -f Dockerfile.hud -t hud-libero-env .
HUD_DOCKER_ENV=hud-libero-env uv run python run.py

hud deploy hosts the env on the platform instead of local Docker.

Docs: docs.hud.ai/v6/advanced/robots