Lower-body backend

Lower-body backend#

A backend is a frozen RL locomotion policy behind the LowerBodyController contract. Every control step it reads a velocity-style command and returns position targets for the joints it owns, the legs and usually the waist. The env routes the base slots of the outer action to the backend, and its joints stay out of the agent’s action space.

`groot_wbc_g1`#

backend	robot	policy
`groot_wbc_g1`	`g1_dex1`	NVIDIA GR00T-WBC (Decoupled WBC, Balance + Walk)

BiGym 2.0 ships one backend, and the benchmark uses it (see Official configuration). Commanded 0.35 m/s, it walks 0.30 m/s.

groot_wbc_g1 runs NVIDIA’s frozen Decoupled WBC policy [1] from GR00T-WholeBodyControl with the weights unmodified.

from bigym.loco import make

make("move_plate")                                   # controller backend groot_wbc_g1
make("move_plate", controller={"backend": "groot_wbc_g1", "cmd_clip": 0.5})

Command bounds#

Each backend declares a typed CommandSpec. GR00T-WBC’s training command ranges are not published, so its adapter declares benchmark clips. A command inside them may still be outside the training distribution.

The env derives the outer height and pitch bounds from this spec unless the config overrides them. Both change action normalization, so the resolved bounds must match the demonstration metadata.

Configuration reference#

bigym.loco.config.ControllerConfig (EnvConfig.controller) configures the controller. controller=None builds a floating base without legs. Common fields, with their official defaults:

field	default	meaning
`backend`	`"groot_wbc_g1"`	backend name
`base_action_mode`	`"lowerbody_cmd"`	base slots carry velocity and height commands
`pitch_command`	`True`	adds the torso-pitch slot (off for six tasks)
`reset_warmup_steps`	`200`	settle steps after reset before the agent engages
`deterministic_reset`	`True`	every reset restores the same controller state, so `reset(seed)` depends only on the seed
`passive_base_tilt`	`True`	unactuated pelvis roll and pitch joints keep the policy’s trained balance dynamics. Adds no action slots and no floating-base proprioception
`cmd_clip` / `wz_clip`	1.0	symmetric clips on `vx`/`vy` and `wz`
`height_cmd_min/max`, `pitch_cmd_min/max`	`None` (from `command_spec`)	override the outer height and pitch slot bounds
`default_height_cmd` / `init_pelvis_z`	0.74 / 0.74	standing height command and reset pelvis height, in m
`model_path`	`None` (the packaged ONNX pair)	comma-separated Balance and Walk policy files

Adding a backend#

Subclass LowerBodyBase. It resolves joint addresses, applies the reset pose, detects falls, clips commands and snapshots replay state.

For replay state, declare STATEFUL, a map from snapshot key to attribute: STATEFUL = {"cmd": "command", "height_cmd": "height_command", ...}. The base class then provides bit-exact get_state and set_state, and set_state raises KeyError when a snapshot misses a key. List every attribute that shapes future targets.

You implement:

__init__: load the policy and set the attributes listed on LowerBodyBase: controlled_joints, command_spec, control_dt, the joint ranges (via build_joint_ranges), the command clips and the latched command.
reset(): put the joints in the reset pose (via apply_pose) and clear the mutable state.
step(): build the observation, run the policy, and return joint position targets in controlled_joints order.
get_base_obs(): base-frame linear velocity, angular velocity and projected gravity. Policies use different frame conventions, so the base class leaves this to you.

bigym/loco/adapters/groot_wbc.py is a complete example. To put the controller in the env with a BackendBinding and select it from your own package, see Building on BiGym 2.0.