Deploy export

Deploy export#

Train and play write robot-agnostic artifacts under each run’s params/ folder. Any hardware runtime that consumes the same contract can load them — no wbc-mjlab install required on the robot computer.

Pipeline#

motion clips ──► data-to-npz ──► train / play
                                         │
                                         ▼
                           params/policy.onnx + config.yaml
                                         │
                                         ▼
                                deploy runtime (robot)

Artifacts#

logs/rsl_rl/<experiment>/<run>/
  params/
    policy.onnx
    config.yaml
    env.yaml
    agent.yaml

config.yaml (schema_version: wbc_tracking_params_v1) lists joint names, observation term order, reference command layout, and PD gains — regenerated from the task config if missing.

Observation layout#

Deploy runtimes must match the exported actor stack — not a generic G1 layout.

Open params/config.yaml from your checkpoint run:

  • actor_observations.<term>.dim — per-term size (J and B already resolved)

  • tracking.actor_observation_names — concatenation order (must match ONNX input)

  • tracking.reference_observation_names — which terms count as reference command

  • tracking.wbc_command_dim — reference motion size for clip playback

Example fragment:

actor_observations:
  ref_base_height: {dim: 1, ...}
  ref_joint_pos: {dim: 29, ...}
tracking:
  actor_observation_names: [ref_base_height, ref_joint_pos, ...]
  wbc_command_dim: 39

Dim rules (before export): Observations. Full schema: Export.

Manual export#

uv run wbc-mjlab-export-tracking-params --task <TaskId> --out /path/to/config.yaml

Example (in-tree reference task):

uv run wbc-mjlab-export-tracking-params --task Wbc-G1 --out /path/to/config.yaml

End-to-end checklist#

# 1. Convert motion library (once per dataset)
uv run wbc-mjlab-data-to-npz --robot g1 --dataset samples --batch-size 8

# 2. Train
uv run wbc-mjlab-train --task Wbc-G1 --dataset samples

# 3. Validate in sim (also writes params/policy.onnx + config.yaml before the viewer)
uv run wbc-mjlab-play --task Wbc-G1 --dataset samples --viewer viser

# 4. Optional: regenerate config.yaml only
uv run wbc-mjlab-export-tracking-params --task Wbc-G1 \
  --out logs/rsl_rl/wbc_g1/<run>/params/config.yaml

# 5. Hand off to a deploy runtime (example: wbc-g1-deploy)
cp logs/rsl_rl/wbc_g1/<run>/params/policy.onnx  /path/to/wbc-g1-deploy/config/policy/
cp logs/rsl_rl/wbc_g1/<run>/params/config.yaml /path/to/wbc-g1-deploy/config/policy/

Play exports ONNX + config.yaml into the checkpoint run’s params/ before the viewer opens. Train checkpoints also keep params/ when the runner exports.

Reference runtime#

wbc-g1-deploy is a reference implementation for one platform (Unitree G1): ONNX inference, config.yaml parsing, and motion clip playback. Use it as a template when building a deploy stack for your robot — the export format is not G1-specific.

See the wbc-g1-deploy README for build and run instructions. Schema details: Export.

Tips#

  • Tasks built with apply_wbc / apply_zest use deploy-style actor obs (no ref_joint_vel) — preferred for sim→real export.

  • SE variants (apply_se_actor) add anchor error + base velocity — export only if your runtime provides the same terms.

  • Validate tracking in sim first: wbc-mjlab-play --viewer viser.