MuJoCo Walker Scaffold
This folder adds a Python + MuJoCo rigid-body simulation track to the existing web-based kinematic simulator.
What is included
- A rigid-body 4-leg linkage model in MJCF (two legs per side, tandem links).
- A fully unconstrained 3D torso base (
freejoint) so roll/yaw/lateral motion emerge from contacts. - One crank motor per side; each motor drives two legs through closed-chain constraints.
- A PD crank controller baseline (phase-driven near/far cranks).
- A CLI runner for quick simulation and optional viewer.
- A CSV trajectory exporter for replay/comparison workflows.
Setup with uv
From this directory:
uv sync
Run the baseline simulation (headless):
uv run walker-run --seconds 10
Evaluate gait quality numerically:
uv run walker-eval --seconds 10
Foot length tuning (middle ground between 5 cm and 30 cm)
In model/walker.xml, lateral foot span is controlled by these capsule geoms:
near_front_foot_lateralnear_rear_foot_lateralfar_front_foot_lateralfar_rear_foot_lateral
Their fromto Y distance is the effective foot length (for example 0.30 means 30 cm).
Practical search strategy:
- Run a coarse sweep from 5 cm to 30 cm.
- Keep only lengths with
fallen=False. - From those, pick the shortest length that still has acceptable forward speed.
- Run a fine sweep around that candidate (for example +/-3 cm).
Use the dedicated sweep command (run from sim/mujoco):
uv run walker-tune-feet --min-cm 5 --max-cm 30 --step-cm 2 --seconds 10
To save results for plotting:
uv run walker-tune-feet --min-cm 5 --max-cm 30 --step-cm 1 --csv ../../artifacts/foot_sweep.csv
Suggested acceptance rule for "middle ground":
fallen=False- pick the smallest stable length first
- use speed/sway as secondary tie-breakers
With the current baseline controller, a good first region to inspect is usually around 8-15 cm, then refine from there.
You can run any simulator command with an overridden foot length (without editing walker.xml) via --foot-cm:
uv run walker-eval --seconds 15 --foot-cm 15
uv run walker-export-web --frame world --seconds 15 --foot-cm 15 --out ../../artifacts/mujoco_trace_15cm.json
Run with interactive MuJoCo viewer:
uv run walker-run --viewer --seconds 30
Render video without opening a GUI:
uv run walker-video --seconds 12 --out ../../artifacts/walker.mp4
Export a CSV trajectory:
uv run walker-export --seconds 15 --out ../../artifacts/walker_traj.csv
Export playback JSON for the browser viewer (true 3D playback + 2D overlay):
uv run walker-export-web --frame world --seconds 15 --out ../../artifacts/mujoco_trace.json
Use --frame world for physically grounded playback in the web app. In world frame, the feet contact the ground plane at z=0 and torso height stays above the floor as expected.
If you export with --frame body, the torso is re-centered to body-local coordinates (torso.z ~= 0), which is useful for relative-linkage debugging but can look like the body is dropped onto the floor in overlay/playback.
In the web app, use "MuJoCo Import" to load that JSON.
You can also load built-in MuJoCo playback presets from the web app "Presets" tab:
-
MuJoCo 10cm trace -
MuJoCo 20cm trace -
MuJoCo 30cm trace -
Default format is
v2(walkersim-mujoco-playback-v2) and is used for true playback in the 3D pane. -
To export the old 2D-only overlay schema, add
--format v1.
RL wrapper
Install RL dependencies:
uv sync --extra rl
Train PPO baseline:
uv run walker-train-ppo --steps 200000
Notes
- The torso now uses an unconstrained freejoint; side tilt is physically simulated, but the default PD crank gait can be less stable and may tumble during some runs.
- The model matches the codebase linkage layout (nearF/nearG/farF/farG) but still needs tuning for stronger forward gait.
- This scaffold is independent from the browser simulator and can be evolved without breaking the JS app.
- Next step is geometry alignment/tuning against the linkage dimensions in the main app.
TODO
- Tune contact/friction/damping to improve stability with unconstrained 3D freejoint dynamics.
- Retune PD gait phase offsets and gains for forward progress under full 3D dynamics.
- Add roll/pitch/yaw summary metrics to
walker-evalfor easier stability regression checks. - Add a short recommended viewer camera preset for inspecting side tilt behavior.