Skip to content

asimovinc/asimov-mjlab

 
 

Repository files navigation

Asimov Locomotion

Fork of mujocolab/mjlab for the Asimov bipedal robot.

Asimov CAD      Asimov Robot


Demo

Sim2Sim in MuJoCo - Multi-directional velocity tracking:

Asimov Sim2Sim Demo


Robot Modeling

12-DOF bipedal (6 per leg): hip_pitch, hip_roll, hip_yaw, knee, ankle_pitch, ankle_roll

Key characteristics:

  • Canted hip pitch axis (45° tilt) - unique kinematics vs standard humanoids
  • Asymmetric left/right joint axes - opposite signs for symmetric motion
  • Narrow stance (11.3cm) - requires conservative velocity limits

Training

Reward Structure

Reward Description
Velocity tracking Track commanded (vx, vy, ωz)
Imitation Match walking reference @ 1.25Hz gait
Alternating feet Enforce bipedal gait pattern
Pose Joint-specific variance (larger for canted hips, tight for ankles)
Air time Encourage dynamic stepping (light robot)
Self-collision Penalize inter-link collisions

Configuration

  • PPO with adaptive LR, 5k iterations
  • Terrain curriculum
  • Network: (256, 256, 128) - sized for 12-DOF

Adaptations vs G1

  • Conservative velocity limits due to narrow 11.3cm stance
  • Higher body_ang_vel penalty (-0.08) for narrow-stance stability
  • Tight ankle pose constraints (limited ROM: ~±30° pitch, ±6° roll)
  • Larger hip variance in pose reward (canted axis couples roll/pitch)
  • Gait clock observation for phase-aware control

Observations (Policy)

Observation Source
base_ang_vel IMU
projected_gravity IMU
velocity_command (vx, vy, ωz)
joint_pos Relative to default pose
joint_vel Joint velocities
previous_actions Action history
gait_clock [cos(φ), sin(φ)] @ 1.25Hz for phase-aware control

Removed: base_lin_vel (no state estimator on real robot)


Sim2Real

  • Physics-based PD gains: KP = J_reflected × ωn² (10Hz), KD = 5.0 Nm·s/rad (hardware max)
  • Zero pose used as default (mechanically stable)

Quick Start

# Install uv if you haven't already
curl -LsSf https://astral.sh/uv/install.sh | sh

# Clone and run
git clone https://github.com/menloresearch/asimov-mjlab.git
cd asimov-mjlab
uv sync

Train Asimov Velocity Policy

uv run train Mjlab-Velocity-Flat-Asimov --env.scene.num-envs 4096

Evaluate Policy

uv run play Mjlab-Velocity-Flat-Asimov --wandb-run-path your-org/mjlab/run-id

License

Licensed under the Apache License, Version 2.0.

Based on mjlab by MuJoCo Lab.

About

Asimov Humanoid Locomotion

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors

Languages

  • Python 99.5%
  • Other 0.5%