| name | gaponet-train |
| description | Run GapONet training. Use when the user invokes /gaponet-train or $gaponet-train, or asks to train a GapONet humanoid policy after data preparation using Isaac Lab, Isaac Sim, rsl_rl, DeepONet, Transformer, MLP, or the unified runner. |
GapONet Train
Run the training stage from the GapONet repository root, identified by scripts/run_gaponet_job.py, check_env.py, and configs/train_default.json.
Workflow
- Ensure data is staged. If
source/sim2real/sim2real/tasks/humanoid_operator/motions/motion_amass/edited_27dof/test.npz is missing, run $gaponet-data or ask for an input .npz.
- Validate the environment before training:
python check_env.py --output-json ./runs/train/env_check_result.json
- If required checks fail for CUDA, Isaac Sim, Isaac Lab, missing scripts, or missing motion data, stop and report the blocking checks.
- Run training through the unified wrapper:
python scripts/run_gaponet_job.py --mode train --config configs/train_default.json --output-dir ./runs/train
- If the user provides overrides such as
max_iterations, num_envs, run_name, device, or task, create a run-local JSON config under runs/train/ by copying configs/train_default.json, apply the overrides, and pass that config to the wrapper.
Defaults
- Training config:
configs/train_default.json
- Default task:
Isaac-Humanoid-Operator-Delta-Action
- Data expected by env config:
source/sim2real/sim2real/tasks/humanoid_operator/motions/motion_amass/edited_27dof/test.npz
- Output directory:
runs/train
- Standard artifacts:
run_manifest.json, training_metrics.json, stdout.log, stderr.log
Completion
Consider training complete only when runs/train/run_manifest.json reports "status": "success" or when a clear blocking failure is reported. Training can be long and Isaac Sim dependent; do not fake completion from command launch alone.
