---

name: export description: "Route AgiBot X1 DH checkpoint-to-JIT export, JIT-to-ONNX conversion, artifact preflight, validation, and backend-aware failure recovery." disable-model-invocation: true metadata: disco-role: operating license: NO_LICENSE

X1 DH policy export

Use this sub-skill only for the registered x1_dh_stand model conversion pipeline and its serialized artifacts. Keep the two stages distinct:

1. Checkpoint → JIT: model_<iteration>.pt → policy_dh.jit.
2. JIT → ONNX: policy_dh.jit → x1_policy.onnx.

The ONNX stage never reopens a training checkpoint, and its --checkpoint argument does not select one.

Route by intent

Intent	Route
Select a runner checkpoint and convert it to JIT	Stay here; use export workflows
Convert an existing JIT policy to ONNX	Stay here; use export workflows
Check paths, checkpoint structure, or a serialized artifact	Stay here; use preflight_export.py and the artifact contract
Produce or resume a training checkpoint	Use training
Run a runner checkpoint interactively in Isaac Gym	Use playback
Run an exported policy in MuJoCo	Use sim2sim after JIT validation

Do not use this route to train, launch Isaac Gym playback, run MuJoCo, or claim robot/simulation safety.

Required backend boundary

Both repository export entry points import humanoid.envs; their parser/helper import chain also requires Isaac Gym Preview 4. That backend is unavailable in the verified construction environment. Therefore full source checkpoint → JIT and JIT → ONNX execution remains:

BLOCKED_REQUIRED_BACKEND: Isaac Gym Preview 4 unavailable

Do not stub Isaac Gym, bypass task registration, or promote a bundled-helper pass to full source-export verification. A compatible source environment must provide the repository's documented Python 3.8, PyTorch 1.13.1/CUDA 11.7, NumPy 1.23.x, editable project install, and verified Isaac Gym Preview 4. ONNX conversion additionally needs a compatible onnx installation.

The bundled preflight deliberately avoids project and Isaac Gym imports. It can perform path checks everywhere and can inspect trusted local artifacts only when the corresponding already-installed model library is available.

Safe operating sequence

1. Obtain an exact run/checkpoint handoff from training. Prefer explicit run names and checkpoint numbers; do not let “latest” select an unrelated artifact.

2. From this sub-skill directory, inspect the helper interface:

   python scripts/preflight_export.py --help
   

3. Follow export workflows for stage-specific preflight, source commands, destination paths, and validation. Deserialize only trusted checkpoints; torch.load may execute pickle payloads.

4. Enforce the fixed X1 contract in artifact contract: input (1, 3102), deterministic output (1, 12), and exact artifact type for the requested stage.

5. Preserve the source script's printed input/output paths, explicit task, run, checkpoint or JIT timestamp, and package/backend versions in the handoff.

6. On failure, use export troubleshooting and keep backend, missing-input, and failed-artifact states distinct.

Non-negotiable artifact rules

• x1_dh_stand is the only verified task contract in this sub-skill.
• Checkpoint → JIT reconstructs ActorCriticDH, strictly loads model_state_dict, and scripts a CPU wrapper containing the actor, state estimator, and long-history encoder.
• JIT → ONNX loads a prior policy_dh.jit; select its timestamp with --load_run. The shared --checkpoint flag is ignored by that converter.
• Source outputs use logs/, not the README's stale singular log/ spelling. JIT and ONNX go to different trees; see artifact contract.
• A CPU load/shape check validates serialization only. It does not verify training-time observation construction, Isaac Gym behavior, MuJoCo behavior, controller integration, or robot safety.