gaponet-inf

name	gaponet-inf
description	Run GapONet inference and evaluation. Use when the user invokes /gaponet-inf or $gaponet-inf, or asks to export a GapONet checkpoint to TorchScript/JIT, run lightweight inference, deploy a policy, evaluate test motion data, or produce inference metrics.

GapONet Inference

Run the inference stage from the GapONet repository root, identified by scripts/run_gaponet_job.py, scripts/rsl_rl/inference_jit.py, and configs/deploy_default.json.

Workflow

1. Ensure data is staged. Prefer configs/deploy_default.json, which should point at the data produced by $gaponet-data.
2. Decide whether export is needed:
   • If the user provides a JIT policy or model/policy.pt exists, run deploy directly.
   • If the user provides a training checkpoint such as model/model_17950.pt, export first, then deploy.
   • If neither exists, inspect runs/train/training_metrics.json or logs/ for the latest checkpoint; if none is available, ask for a checkpoint or JIT model path.
3. Export a training checkpoint to JIT when needed:

python scripts/run_gaponet_job.py --mode export --checkpoint <checkpoint.pt> --output-dir ./runs/inf/export

4. Run lightweight inference and evaluation:

python scripts/run_gaponet_job.py --mode deploy --config configs/deploy_default.json --output-dir ./runs/inf/deploy

5. If using the JIT model from the export step, pass it explicitly:

python scripts/run_gaponet_job.py --mode deploy --checkpoint ./runs/inf/export/policy.pt --input-data ./source/sim2real/sim2real/tasks/humanoid_operator/motions/motion_amass/edited_27dof/test.npz --output-dir ./runs/inf/deploy

Notes

• Export requires Isaac Sim because it creates the task environment to infer policy dimensions.
• Deploy is lightweight and does not require Isaac Sim, but it does require PyTorch, a JIT model, and test .npz data.
• Standard deploy artifacts are run_manifest.json, eval_metrics.json, stdout.log, and stderr.log.

Completion

Consider inference complete only when runs/inf/deploy/run_manifest.json reports "status": "success" and eval_metrics.json exists. Report payload-grouped metrics when present.