Locomotion robots#
The bundled locomotion presets use the generic Robot runtime. Each model has
its own page with an individual render, usage example and configuration summary.
They can be selected independently of a task through the robot catalog:
Catalog ID |
Configuration class |
Controlled joints |
|---|---|---|
|
12 |
|
|
12 |
|
|
29 |
|
|
27 |
|
|
12 |
|
|
14 |
embodichain preview-robot --robot UnitreeGo2
Each configuration owns the simulation asset, floating base, default standing
pose, locomotion control group and effective PD drive defaults previously
defined by its flat-velocity task. The same model configuration serves Default
and Newton deployments. G1 and H1_2 identify the specific bundled models above;
they do not select other hardware revisions implicitly.
Task YAML selects robot.class_type and may override joint_drive_props.
Tasks continue to own action scaling, observation layout, commands, rewards,
randomization and illegal-contact rules. Contact sensors remain environment
declarations. The model profiles are simulation presets, not certified hardware
calibrations or implementations of Isaac Lab’s learned or BAM actuators.
The backend-free profile reader is shared with pure Torch task contracts.
Source-ordered asset.init_qpos and the ordered
robot.joint_names / robot.default_joint_position pair have distinct roles;
keep their ordering when changing a model. Foot and IMU reference points are
stored alongside the model. The presets retain build_pk_chain=False and
provide no limb IK solver.
See the API reference for the public configuration classes and model-profile reader.
Render Settings#
The individual previews use DexSim OfflineRT (rt) with shared lighting and
render settings: 1 sample per pixel, DLSS denoising, tone mapping disabled and
exposure 1. Each model uses its default joint pose and is framed independently,
so the images do not show a common physical scale. RGB captures have no
additional gamma or color correction.