Robots#

Choose a robot with --robot-name when building a training configuration. These pages describe the default models registered in protomotions/robot_configs/factory.py, not every asset variant or physical robot that can be integrated into ProtoMotions.

Default Models#

Model / CLI name

Bodies

Joint DOFs

Forward

Observation anchor

Typical input

Unitree G1 / g1

33

29

+X

torso_link

G1-retargeted motion

Unitree H1-2 / h1_2

29

27

+X

pelvis (root)

H1-2-retargeted motion

SOMA23 / soma23

23

66

-Y

Hips (root)

Converted SOMA motion

Anatomical SOMA / anatomic_soma

53

52

+X

Chest

Retargeted SOMA motion

SMPL Humanoid / smpl

24

69

+X

Pelvis (root)

Converted SMPL motion

SMPL-X Humanoid / smplx

52

153

+X

Pelvis (root)

Converted SMPL-X motion

AMP Humanoid / amp

15

36

+X

pelvis (root)

Motion matching the AMP humanoid

Counts refer to the configured MJCF, exclude the world body and floating-base DOFs, and include fixed bodies and intermediate hinge carriers. Anatomical SOMA has 23 semantic parts plus 30 carriers. All models use Z up; positions are in meters and scalar joint angles are in radians. A root anchor means anchor_body_name=None, which resolves to body index zero.

Configuration And Compatibility#

A robot config supplies the asset, semantic body mappings, control parameters and backend-specific simulation settings. An experiment may override these defaults before saving resolved_configs.json. Resume uses that saved config; inference starts from the saved config and applies inference overrides. The pages here do not override the contract of an existing checkpoint.

Use motion files and checkpoints prepared for the same articulation, joint order, body frames and observation layout. Matching DOF counts alone does not establish compatibility. The observation anchor is distinct from the physical root and from the bodies selected for target tracking.

PD gains and effort limits are resolved from the MJCF and the config’s control.override_control_info. The action-to-target mapping belongs to the experiment. Changing gains, action scaling or simulation frequency can change an existing policy’s behavior; a backend config is not evidence of sim-to-sim transfer quality.

Assets And Workflows#

Asset paths on each page are relative to protomotions/data/assets/. Use a source checkout for the linked example experiments and data tools. Git LFS payloads must be materialized where required. SMPL-family assets and body models have separate licensing and installation requirements; see their pages.