praxis-sim

The MuJoCo simulation of the Praxis-1 single-arm rig (a Franka arm with a Robotiq 2F-85 gripper, an exterior and a wrist camera, 15 Hz control) and its fixed evaluation suites (droid, droid_guard, hard), and the bimanual YAM rig (two YAM arms, a top camera and a camera per wrist, 11 tasks: yam_scene, yam_env, yam_tasks, yam_scripted), packaged on their own: MuJoCo, Gymnasium and numpy, no torch. The model loop (praxis.sim.loop), the video writer and the live viewer stay in the praxis package, which runs the simulation through this one.

pip install praxis-sim
# or, from a checkout of this repository (the `sim` extra of praxis pulls the package in too):
pip install -e packages/praxis-sim

Python 3.10 or newer (the praxis package itself needs 3.12). The package has no torch and imports no core praxis module; tests/test_isolation.py checks that, and the tests run in an environment that holds only the package:

python -m venv /tmp/praxis-sim-venv && /tmp/praxis-sim-venv/bin/pip install "./packages/praxis-sim[dev]"
MUJOCO_GL=egl /tmp/praxis-sim-venv/bin/pytest packages/praxis-sim/tests

The robot models (MuJoCo Menagerie, pinned commit) download on first use into a cache directory and are checked by SHA-256 (praxis_sim.assets).

Gymnasium environments

praxis_sim.gym_env.PraxisSimGymEnv is one task as a Gymnasium environment: the reset seed is the scene seed, the observation holds the two camera views (pixels/exterior, pixels/wrist, 512 x 288 uint8), the pose of the Franka flange as [x, y, z, roll, pitch, yaw, gripper] (agent_pos, the DROID convention) and the scene seed; the action is the same seven values as an absolute target. create_praxis_sim_envs(task, n_envs, ...) builds the {suite: {task_index: VectorEnv}} layout of a suite, the same episodes as scripts/run_sim_eval.py of the praxis repository.

Comparing runs

python -m praxis_sim.compare <lerobot eval_info.json> <praxis runner output directory> pairs the episodes of a LeRobot evaluation with those of the praxis runner by task and scene seed and reports wins, losses, the exact McNemar p and a bootstrap interval of the mean difference.

Layout