Skip to main content
The Unitree H1 is one of the most accessible bipeds in research right now, with a MuJoCo Menagerie model that includes correct mass distribution, joint limits, and a tuned default standing keyframe. This guide does not train a walking policy — that’s a research project on its own. What it does is everything you need around a walking policy or before you have one: scene, balanced PD hold-pose, full sensor wiring, viewer launch, ROS2 bridging on request.
Recommended setup: MuJoCo 3.8.1+, Python 3.10+. Drift can clone Menagerie for you on the first run if it isn’t on disk yet.

​
The prompt

set up a Unitree H1 from mujoco_menagerie standing on flat ground,
write a PD controller that holds the default standing keyframe on
all 19 joints with stiff hips and ankles, run 5 seconds headless,
and log IMU accel + gyro, base position + orientation, and all 19
joint angles + torques to sensor_log.csv
From this prompt — and a tuning follow-up or two, since humanoid PD stability is gain-sensitive — Drift:
  • Asks where mujoco_menagerie lives, or clones it to a path you confirm
  • Composes h1_scene.xml that <include>s the Menagerie H1 and adds a 10×10m floor
  • Spawns the H1 at the default keyframe’s qpos so the feet rest cleanly on the ground
  • Writes a run_sim.py that:
    • Loads the H1 keyframe pose as the joint-target reference
    • Applies a per-joint PD controller — an initial gain set with stiffer hips and ankles, lighter arms (the common starting point for hold-pose stability; expect to retune)
    • Steps at 1 ms timestep with implicitfast integrator
    • Logs sensors at 200 Hz
  • Runs headless, prints the final base height and orientation drift, and notes any contact issues
Static stability of a 19-DOF biped on PD is sensitive to gains. If the H1 collapses on the first run, that’s normal — see the gotchas at the bottom for the usual fixes.

​
What ends up on disk

your-project/
├── h1_scene.xml            # floor + Menagerie H1 include
├── run_sim.py              # full-body PD hold-pose + sensor logger
├── sensor_log.csv          # 1000 rows: IMU, base pose, 19 joints (pos + torque)
└── DRIFT.md                # scene summary, gains used, what to iterate

​
What the viewer shows

open h1_scene.xml in the MuJoCo viewer
If the gains are well-tuned, the H1 stands at the default pose — feet flat on the floor, knees slightly bent, arms by the sides — with IMU orientation steady within a few mrad. The torque trace shows the static hold-pose torques on the hips and ankles where most of the gravity load lives. Push the robot in the viewer (drag with the right mouse button): a well-tuned controller resists and settles back; an under-tuned one folds. If it folds, the gotchas below cover the fix.

​
Iterating from here

This is where the real work starts: 
add a small perturbation: apply a 30N forward push to the H1's
torso at t=2s for 50ms, see if it recovers
add stairs in front of the H1 and have it step forward onto the
first step (lift right foot, plant, shift weight)
load an MPC walking controller from /path/to/h1_mpc and bridge it
to the joint targets — log the contact schedule and CoM trajectory
add a head-mounted RGB camera and stream 30fps frames
write a ROS2 bridge: publish /h1/joint_states, /h1/imu, /h1/base_pose
and subscribe to /h1/joint_cmd for external controllers
load /path/to/h1_walking_policy.onnx and step it every 20ms,
feeding the standard humanoid observation (joint pos + vel + IMU +
base height), outputting target joint positions

​
Honest scope

Things this guide does:
  • Compose the scene around the Menagerie H1 model
  • Write a full-body PD hold-pose controller (stiff hips/ankles, lighter arms)
  • Wire IMU + base pose + 19-joint sensors and log them
  • Open the viewer / run headless
Things this guide does not do (and will not pretend to):
  • Bipedal walking — even “in place” walking is non-trivial; you need a controller (MPC, learned policy, or a reference trajectory tracker), not Drift
  • Push recovery beyond what hold-pose PD naturally absorbs
  • Whole-body control with operational-space inverse dynamics — Drift can scaffold the controller skeleton but the math is yours
  • Sim-to-real — H1 sim parameters are good but not identical to hardware
The honest framing: this guide gets you to a stable platform you can plug your controller into. The “wow” moments — push recovery, walking, getting up — come from the controller you add, not from this scene file.

​
Common gotchas

Almost always one of:
  • Keyframe wasn’t loaded — mj_resetDataKeyframe call missing in the run script.
  • PD gains too low for hip flexion / knees. The Menagerie default ranges can vary by revision; ask Drift "the H1 is collapsing — tune the hip and knee gains for static stability".
  • Floor friction is too low (humanoids slip more than quadrupeds). condim=3 on the floor, friction ≥ 1.0.
Expected with pure PD on a model with no integral term. Either:
  • Accept the drift (often fine for testing)
  • Switch to PID by adding small Ki integral terms — ask Drift to do this
  • Use the Mujoco-MPC reference controller if you have it
Stiff PD on a 19-DOF humanoid will saturate some joints in the first few timesteps as the controller “catches” the system. If the warnings stop after ~0.1s, the system is fine. If they persist, lower your kp or check the keyframe is actually loaded as the target (not zero).
The default Menagerie spawn is at qpos[3:7] = (1, 0, 0, 0) (identity quaternion, facing +X). If you want it facing a different direction, ask:
spawn the H1 facing -Y instead of +X
and Drift will compute the right quaternion.

​
Next steps

Quadruped Go2

Quadruped showcase — easier locomotion baseline

Dual-arm Aloha

Manipulation showcase

Building a MuJoCo Scene

The base MuJoCo guide

Custom Skills

Capture your humanoid stack’s conventions (gain rules, contact tuning) as a skill so every prompt uses them