Robot Model (v1)

The simulation model of our course robot, for MuJoCo and mjlab: a 12-joint biped with 6 joints per leg. It includes the robot geometry and a first set of actuator parameters.

Download ASMR2026 Humanoid v1 (ZIP, 10 MB)

What’s inside

File Description
mjcf/asmr2026_humanoid.xml MuJoCo model with position actuators and sensors
mjcf/scene.xml The MuJoCo model on a ground plane
mjlab/asmr2026_humanoid_mjlab.xml Same geometry without actuators; mjlab adds them from the config below
mjlab/asmr2026_humanoid_constants.py Actuator parameters and the mjlab robot config (get_asmr2026_humanoid_cfg())
meshes/ Visual meshes (STL), in the ZIP only
README.md Short overview of the package

Quick start

Unzip the package and open the model in the MuJoCo viewer:

pip install mujoco
python -m mujoco.viewer --mjcf=asmr2026-humanoid-v1/mjcf/scene.xml

For mjlab, use get_asmr2026_humanoid_cfg() from asmr2026_humanoid_constants.py as the robot in your task config. See the mjlab page for mjlab itself.

For Isaac Sim / Isaac Lab, convert the model to USD yourself. See Isaac Sim and Isaac Lab.

Frames and joints

  • Frame: +X forward, +Y left, +Z up.
  • Joints, per leg from hip to foot: hip_roll, hip_yaw, hip_pitch, knee_pitch, ankle_pitch, ankle_roll, named leg_left_*_joint and leg_right_*_joint.
  • IMU: the imu site.
  • Feet: the sole boxes left_foot_collision and right_foot_collision, with sites left_foot and right_foot.

Actuator parameters

Joints Actuator Kp Kd Torque limit Armature Friction (static / viscous)
hip roll, hip yaw, hip pitch, knee pitch 6512 90.01 5.730 18.09 N·m 0.0228 kg·m² 0.377 N·m / 0.090 N·m·s/rad
ankle pitch, ankle roll 5010 21.95 0.489 22.57 N·m 0.00556 kg·m² 0.317 N·m / 0.030 N·m·s/rad

Armature and friction were fitted from bench tests on a single actuator of each type. Kp and Kd are derived from the armature with a 10 Hz natural frequency and a damping ratio of 2.0 (hip and knee) or 0.7 (ankle). The torque limit comes from the motor driver’s 18 A current limit.

These parameters may not be accurate for your robot. We recommend running your own experiments on your real actuators and building your own actuator model. For example, the ANYmal team at ETH Zurich’s Robotic Systems Lab trained a neural network that maps actuator commands to output torque from real measurements: Hwangbo et al., “Learning agile and dynamic motor skills for legged robots”, Science Robotics, 2019.

See Actuator Modeling for methods you can use.

Versions

Version Date Notes
v1 2026-10-16 First release

Copyright © 2026 Seunghoon Kang, Soft Robotics and Bionics Laboratory (SRBL), Department of Mechanical Engineering, Seoul National University. All rights reserved. Provided for use in the 2026-2 Actuation and Sensing Mechanism Design for Robots course.