Implementation:Haosulab ManiSkill UnitreeG1UpperBody

Knowledge Sources	Haosulab_ManiSkill
Domains	Robotics, Simulation, Humanoid_Control
Last Updated	2026-02-15 08:00 GMT

Overview

Concrete tool for simulating the Unitree G1 humanoid robot upper body (with fixed legs) in ManiSkill environments.

Description

The UnitreeG1UpperBody class models the Unitree G1 humanoid with control limited to the torso rotation and both arms. The legs are fixed (fix_root_link=True). It has 25 active DOF: 1 torso joint, 4 shoulder joints per arm (pitch, roll, yaw), 2 elbow joints per arm (pitch, roll), and 7 finger joints per hand (zero, one, two, three, four, five, six). Custom collision groups are set to disable finger-to-finger, palm, and torso collisions for simulation stability. The file also defines two variants: UnitreeG1UpperBodyWithHeadCamera (adds a head-mounted camera) and UnitreeG1UpperBodyRightArm (controls only the right arm, with passive joints for the rest).

uid: unitree_g1_simplified_upper_body

URDF path: {PACKAGE_ASSET_DIR}/robots/g1_humanoid/g1_simplified_upper_body.urdf

Supported control modes: pd_joint_delta_pos, pd_joint_pos

Usage

Use UnitreeG1UpperBody for bimanual manipulation tasks requiring both arms and dexterous fingers. The fixed legs simplify the task to upper-body manipulation only. The right-arm-only variant is useful for single-arm grasping tasks. The head camera variant supports visual manipulation tasks.

Code Reference

Source Location

• Repository: Haosulab_ManiSkill
• File: mani_skill/agents/robots/unitree_g1/g1_upper_body.py

Signature

@register_agent()
class UnitreeG1UpperBody(BaseAgent):
    uid = "unitree_g1_simplified_upper_body"
    urdf_path = f"{PACKAGE_ASSET_DIR}/robots/g1_humanoid/g1_simplified_upper_body.urdf"
    fix_root_link = True
    load_multiple_collisions = False
    body_joints = ["torso_joint", "left_shoulder_pitch_joint", ...]  # 25 joints
    body_stiffness = 1e3
    body_damping = 1e2
    body_force_limit = 100
    keyframes = dict(standing=Keyframe(pose=sapien.Pose(p=[0, 0, 0.755]), qpos=np.array([0.0]*25)))

@register_agent()
class UnitreeG1UpperBodyWithHeadCamera(UnitreeG1UpperBody):
    uid = "unitree_g1_simplified_upper_body_with_head_camera"

@register_agent()
class UnitreeG1UpperBodyRightArm(UnitreeG1UpperBody):
    uid = "unitree_g1_simplified_upper_body_right_arm"

Import

from mani_skill.agents.robots.unitree_g1.g1_upper_body import UnitreeG1UpperBody

I/O Contract

Inputs

Name	Type	Required	Description
(inherited from BaseAgent)

Outputs

Name	Type	Description
robot agent	UnitreeG1UpperBody	Configured upper-body humanoid with torso, dual-arm, and finger controllers

Usage Examples

Creating Environment with Robot

import gymnasium as gym
import mani_skill.envs

env = gym.make("PickCube-v1", robot_uids="unitree_g1_simplified_upper_body")
obs, info = env.reset()

Related Pages

• Environment:Haosulab_ManiSkill_Python_SAPIEN_Core