Skip to main content

Quick start

This page runs everything in simulation (sim2sim). For hardware see Real robots.

1. Smoke test the base loop​

The repository ships a zero-action test policy. It only checks that the build, robot model and viewer work; the robot will not balance.

boltrt --robot g1 --policy examples/base-policy.yaml --backend mujoco

2. Track a motion with a trained policy​

Download a checkpoint, for example mimic-lite/v1_1, from the checkpoint folder. A deploy checkpoint is just policy.onnx + policy.yaml.

boltrt --task tracking --robot g1 --backend mujoco \
--policy mimic-lite/v1_1/policy.yaml \
--motion motions/walk1_subject1.npz

The motion must be retargeted for the same robot. Without --motion, the policy tracks a static reference at its default pose.

3. Drive it​

With the terminal focused:

KeyAction
IMove toward the policy's default pose
]Enter the policy; the reference waits at frame 0
SpacePlay / pause the reference motion
OHold the current pose
QQuit

With the MuJoCo window focused, its native shortcuts apply: Space pauses physics, Backspace resets physics, F1 shows help. Motion, policy and physics are controlled independently; see Tracking task.

4. Batch and measure​

boltrt --task tracking --robot g1 --backend mujoco \
--policy mimic-lite/v1_1/policy.yaml --motion motions/walk1_subject1.npz \
--headless --start --run-once --no-keyboard \
--record outputs/run.jsonl --timings outputs/latency.csv

--headless drops the window and real-time pacing; --record writes state, observations, actions and commands per tick; --timings writes per-tick latency. --validate checks configs, assets and ONNX bindings without running.