wbc¶

wbc and wbc_gait run NVIDIA's GR00T Whole-Body-Control ONNX locomotion controllers for the Unitree G1 in process.

By the end of this page you can make a simulated Unitree G1 walk from a [vx, vy, omega] command with the GR00T Whole-Body-Control (SONIC) controllers, know which checkpoint files it expects, and know why the MuJoCo backend installs a torque shim for it.

pip install 'strands-robots[wbc]'    # onnxruntime + pyyaml + huggingface_hub; no weights bundled

What it is¶

WBCPolicy ports the non-gait reference runner from NVlabs/GR00T-WholeBodyControl (run_mujoco_gear_wbc.py): an 86-wide observation, a 7-wide command block, and two ONNX policies, a balance policy.onnx and a walk_policy.onnx selected by commanded velocity. It drives the G1's 15 leg and waist joints and holds the arms at their nominal defaults. requires_images is False; the controller reads joint state and base IMU only. Layer an arm policy on top with CompositePolicy (legs and waist from wbc, arms from a manipulation policy).

WBCGaitPolicy (wbc_gait) ports the gait-clock variant (run_mujoco_gear_wbc_gait.py): a 95-wide observation with a step-frequency command slot and a two-element left/right foot phase clock, and a single ONNX policy whose input is [batch, 570]. Everything else (SONIC PD gains, name-resolved joint map, checkpoint resolution) is inherited.

from strands_robots.policies import create_policy

policy = create_policy("wbc", checkpoint="./GR00T-WholeBodyControl", target_velocity=[0.5, 0.0, 0.0])
policy = create_policy("sonic", checkpoint="./GR00T-WholeBodyControl")             # same provider
gait = create_policy("wbc_gait", checkpoint="./gait-ckpt", gait_frequency=1.5)

Constructor keywords¶

wbc:

keyword	type	default
`checkpoint`	`str \\| None`	`None`
`config`	`str \\| dict[str, Any] \\| WBCConfig \\| None`	`None`
`walk`	`bool`	`True`
`target_velocity`	`list[float] \\| None`	`None`
`allow_missing_models`	`bool`	`False`
`**kwargs`		unknown keywords are ignored

wbc_gait:

keyword	type	default
`checkpoint`	`str \\| None`	`None`
`config`	`str \\| dict[str, Any] \\| WBCConfig \\| None`	`None`
`target_velocity`	`list[float] \\| None`	`None`
`gait_frequency`	`float \\| None`	`None`
`allow_missing_models`	`bool`	`False`
`**kwargs`		unknown keywords are ignored

checkpoint is a directory holding the ONNX files and an optional config.json, a path to the main .onnx, or a HuggingFace model id. The loader accepts the official names GR00T-WholeBodyControl-Balance.onnx and -Walk.onnx verbatim. When a G1 checkpoint ships ONNX only, the SONIC gains and default angles for 15 actuators are applied. walk is a strict boolean; "false" is refused, not read as truthy. target_velocity in the constructor is the default command for paths that forward constructor kwargs only, such as the mesh tell(); the per-call keyword overrides it.

The repo nvidia/GEAR-SONIC ships the SONIC VLA inference stack (model_encoder.onnx, planner_sonic.onnx, ...), not these controllers; pointing checkpoint at it is refused with the reason. Its decoder is what wbc_latent runs, for a VLA that predicts SONIC motion tokens.

Goals¶

keyword	shape	meaning
`target_velocity`	`[vx, vy, omega]`	m/s, m/s, rad/s in the base frame; three components required here
`target_orientation`	`[roll, pitch, yaw]`	optional torso orientation command

Run it¶

Needs the extra and a downloaded checkpoint. run_policy on MuJoCo detects a WBCPolicy anywhere in the policy tree and installs WBCTorqueController, SONIC's per-joint PD law on the compiled model; without it the scene's position servos win, the robot falls in a fraction of a second and the rollout reports success. The Isaac and Newton backends cannot install the shim and refuse to start the rollout unless you pass wbc_install_torque_control=False against a torque-actuated scene.

from strands_robots.simulation import create_simulation

sim = create_simulation("mujoco")
sim.create_world()
sim.add_robot("g1")
result = sim.run_policy(
    robot_name="g1",
    policy_provider="wbc",
    policy_config={"checkpoint": "./GR00T-WholeBodyControl"},
    policy_kwargs={"target_velocity": [0.5, 0.0, 0.0]},
    duration=10.0,
    control_frequency=50.0,
)
print(result["status"])

Robot("g1") on hardware takes the same policy_provider and policy_config.

holosoma is the second whole-body family for the same robot: Amazon FAR's Apache-2.0 controllers, the same 29-joint table and the same torque shim, a 100-wide observation with a gait clock in place of the height command.

Limits¶

Unitree G1 only. The joint map resolves the G1's 29 names by name inside the caller's state keys; a different humanoid fails that membership check.
wbc has no reference-pose input. It cannot track a kimodo motion; that pairing is protomotions.
wbc_gait needs a gait-clock checkpoint ([batch, 570] input, [batch, 15] output). The shipped Balance and Walk weights are the non-gait family (516-wide) and are rejected by shape.
Weights are fetched under the NVIDIA Open Model License; nothing is bundled.

Edit page