Getting started

Install BiGym 2.0 first.

Make and reset

make(task) builds the task’s official configuration: the Unitree G1 with Dex1-1 two-finger grippers (g1_dex1), the GR00T-WBC lower body (groot_wbc_g1), three 84×84 cameras, and the task’s episode budget and success hold.

from bigym.loco import make

env = make("move_plate")
timestep = env.reset(seed=620000)
env.config                    # the resolved EnvConfig

reset runs a warmup before it returns: the lower-body controller steps with the arms held while the robot settles into its stance. Your policy takes over after the warmup. reset_warmup_steps defaults to 200 control steps (4 s). Never set it to 0. The demonstrations start after the warmup, so evaluation has to warm up too.

One step

make returns a dm_env-style wrapper. step(action) advances one 50 Hz control step and returns an ExtendedTimeStep with rgb_obs, low_dim_obs, reward and step_type.

import numpy as np

spec = env.action_spec()                        # shape (20,), float32
action = np.zeros(spec.shape, dtype=spec.dtype)
timestep = env.step(action)
timestep.rgb_obs.shape        # (3, 3, 84, 84): head, right_wrist, left_wrist
timestep.low_dim_obs.shape    # (50,)
timestep.step_type.name       # 'MID'

Actions are normalized to [-1, 1] per slot, and the env maps each slot onto its range. A zero action commands zero base velocity and the middle of every other range.

What am I controlling?

env.wholebody_action_layout()
# {'action_dim': 20, 'base_dim': 4, 'limb_names': (...14 arm joints...),
#  'gripper_count': 2, 'action_low': ..., 'action_high': ..., ...}

env.low_dim_component_slices()
# {'proprioception': (0, 44), 'proprioception_grippers': (44, 46),
#  'proprioception_floating_base': (46, 50)}

The four base slots are [vx, vy, height, wz]: forward and lateral velocity in m/s, absolute pelvis height in m and yaw rate in rad/s. The 14 arm slots are absolute joint position targets. The 2 gripper slots open (0) or close (1) each hand.

Those are the move_plate values. The 14 tasks with the torso-pitch command add a fifth base slot and have a 21-dim action and a 56-dim low-dimensional observation. Read the layout from the env. Official configuration has the full tables.

Overrides

Keyword overrides or an EnvConfig change any setting. The env records what departs from the official configuration, and evaluation reports a run as unofficial when a departure affects results:

env = make("move_plate", camera_keys=("head",), controller={"cmd_clip": 0.5})
env.config_overrides       # {'controller.cmd_clip': 0.5, 'camera_keys': ('head',)}

env = make("move_plate", env.config.override(camera_shape=(128, 128)))

Demonstrations

from bigym.loco import make_gym

env = make_gym("move_plate")
demos = env.get_demos(60)     # downloads move_plate's demos on first use
demo = demos[0]
demo["obs"]["rgb"].shape      # (T + 1, 3, 3, 84, 84)
demo["action"].shape          # (T, 20), in env.action_space

make_gym returns a gymnasium env with the same task. bigym-download --all fetches every task ahead of time. See Demonstrations.

Check the install

MUJOCO_GL=egl uv run python examples/loco_random_agent.py

It ends with one line per seed, such as seed 620000: total reward 0.000, last step MID.