Notebook Workflow

This directory splits each model into two independent Notebook files that can be executed separately:

• Normal training: Start training with normal data and a base model, then conduct a closed loop evaluation.
• Protection training: Reproduction of the training recipe that yields protection results, followed by another closed loop evaluation.

The two modes use different output directories, and the checkpoint, logs, videos, or evaluation results of each other will not be overwritten. The training loops in the Notebook are actual executed Python cells. For long training, a compact progress bar will be displayed every 100 steps in the current cell; step metrics are written to notebook_train_metrics.jsonl, and the start time, end time, total duration, and checkpoint are written to training_run_summary.json at the beginning of each training. No line of Notebook output is generated for each step.

Regular Training

Protection Training

Implementation Recommendations

1. First execute code/rocm/activate_lerobot_rocm.sh, then run code/rocm/verify_lerobot_rocm_env.py; proceed only when ENVIRONMENT_OK is output.
2. For the first learning, run the environment check and smoke cell in the ordinary Notebook first.
3. After confirming that data, model, and device are normal, execute the long training cell.
4. After training, run the Full 14 evaluation cell at the end of the Notebook to check the success rate, stage statistics, video, and action trace of this training.
5. When you want to reproduce the protection results, open the protected Notebook separately; do not modify the path in the ordinary Notebook to replace the protection recipe.

The two AMD devices should use the same set of Python 3.10, ROCm 7.13, PyTorch 2.11, and legacy LeRobot commit.395 workstations. The W7900 cloud instance must run the same verification script after SSH is restored. For detailed version details, see [code/rocm/ROCM_LEROBOT_ENV.md](…/…/…/…/…/16-Special Team Learning/04-AMD-ROCm Policy Replication Topic/code/rocm/ROCM_LEROBOT_ENV.md); for the dependency list, see [requirements-rocm713.txt](…/…/…/…/…/16-Special Team Learning/04-AMD-ROCm Policy Replication Topic/code/rocm/requirements-rocm713.txt).

The results of each notebook are saved in the corresponding outputs/ordinary/<model> or outputs/protected/<model> under the special topic list. The evaluation cell only accepts TRAINED_POLICY_PATH generated after training the current notebook. If there is no checkpoint, an error will be reported, and the model will not revert to historical or pre-trained weights. The training time depends on the model, steps, and GPU; the progress bar only refreshes periodically, and detailed metrics are based on JSONL and training_run_summary.json.

Full 14 Evaluation and Visualization

The regular and Protected versions of each model contain an independent Full 14-episode closed-loop evaluation cell. After execution, it will:

• Run the closed loop with 14 fixed seeds in real time, and generate the complete result.jsonl;
• Keep the video of the first physical success round and the first physical failure round;
• Save the action traces for these two rounds;
• Play the success/failure videos in subsequent cells, and draw the action sequence diagram;
• Summarize the results using physical_success=x/14.

These two case videos and sequence diagrams come from this evaluation, not the pre-set static results in the Notebook. The evaluation uses RENDER_EVAL or the rendering settings in the Full evaluation cell; when no desktop environment is available, the Notebook attempts to launch Xvfb.