Notebook Practical Operation Entry
This directory contains the accompanying Notebook for the AMD ROCm policy replication project. Sections 01–06 focus on the environment, metrics, and diagnostics; sections 07–13 complete the full execution chain from keyboard data collection, formal training, MuJoCo closed-loop, pi0 strict-input to Pi0.5 EEF-delta/chunk alignment diagnosis; sections 14–16 follow the “one Notebook per model” approach, placing training, long-term training logs, strict evaluations, videos, and conclusions in the same file. The Markdown chapters explain the concepts, evaluation criteria, and experimental conclusions, while the Notebook executes the code line by line, generates configurations, starts tasks, and organizes result tables.
07–16 The output was executed line by line and saved. 07–12 Used the AMD teaching environment; 13 Recomputed tables and images in a isolated Jupyter environment using the same AMD experiment summaries. 14–16 Used the completed model to reconstruct evidence and generate an end-to-end learning interface. Interactive keyboard collection, formal long training, and batch closed-loop processes are still controlled by the RUN_* switch. The repository includes path audits, real data summaries, completed training progress, strict success rates, key frames, and four-view videos. These tasks do not start automatically when readers open the Notebook.
It is recommended to start Jupyter from the topic root directory:
bashcd /path/to/04-AMD-ROCm策略复刻专题
jupyter lab
If it is running on your own AMD device or remote server, set it according to the actual situation first:
bashexport PROJECT_ROOT=/path/to/every-embodied/mujoco_pnp
export DATA_ROOT=/path/to/datasets/every_embodied
export OUTPUT_ROOT=/path/to/outputs
export MODEL_ROOT="$PROJECT_ROOT/ckpt"
| Notebook | Corresponding Chapter | Main Purpose |
|---|---|---|
| 01_device_env_check.ipynb | 01 Device and Environment Verification | Check ROCm, PyTorch, memory, temperature, and directory planning |
| 02_physical_success_review.ipynb | 02 Physical Success Evaluation | Understand physical_success, review success/failure key frames |
| 03_act_dagger_diagnostics.ipynb | 03 ACT Diagnosis | View ACT progress curve, organize DAgger evaluation commands |
| 04_smolvla_weighted_sampling.ipynb | 04 SmolVLA Weighted Sampling | Compare success rates of red/blue cups, regenerate charts |
| 05_pi0_smoke_gate.ipynb | 05 pi_0 Training Gatekeeping | Check gated permissions, 1-step smoke, and training command templates |
| 06_rocm_debug_playbook.ipynb | 06 Troubleshooting Review | Organize issues by “phenomenon, evidence, cause, fix, verification” |
| 07_data_collection_and_audit.ipynb | 07 Data Collection | Keyboard collection, strict threshold storage, four-view video recording, audit of 20 existing data measurements |
| 08_act_training_rocm.ipynb | 08 ACT Training | Generate smoke/full configurations, start training, view 5000-step history progress and logs |
| 09_smolvla_training_rocm.ipynb | 09 SmolVLA Training | Complete basic weight loading, training configuration, history progress, and instruction division checks |
| 10_pi0_training_rocm.ipynb | 10 pi_0 Training | Check gated permissions, complete smoke/full configurations, and review historical training results |
| 11_mujoco_closed_loop_deploy.ipynb | 11 Closed Loop Deployment | Run checkpoint, save JSONL, strict success rate, key frames, and four-view video |
| 12_pi0_strict_input_end_to_end.ipynb | 12 pi0 Strict Input | Compare raw/learned heads, fixed/random environments, reproduction of final strict judgment |
| 13_pi05_random_position_eef_delta.ipynb | 13 Pi0.5 EEF-delta | Audit phase direction, coherent recovery, parquet row order, prefix DAgger, action chunk, and new seed strict results |
| 14_smolvla_end_to_end.ipynb | SmolVLA End-to-End | Complete zero-training preview, long training logs, strict evaluation, 57/60 results, and success/failure videos in one Notebook |
| 15_pi0_end_to_end.ipynb | Pi0 End-to-End | Complete Pi0 permission check, protected training, long training logs, 12/14 evaluations, failure analysis, and diagnosis videos in one Notebook |
| 16_act_end_to_end.ipynb | ACT End-to-End | Complete ACT smoke, protected training recipe, long training logs, and strict evaluation in one Notebook; current auditable protected candidate is 15/30, with historical 17/30 as reference |
When learning from scratch for the first time, the recommended order is 01 → 14. First, let the learner see the correct behavior of SmolVLA, long training logs, and the strict evaluation criteria, then proceed to 07 → 09 → 11 to collect, train, and evaluate on their own. If you want to understand failure diagnostics, read 15 → 16 → 02/03/05/06. If there is already data and a checkpoint, you can directly enter the corresponding model from the end-to-end Notebook in 14–16, and then return to 07–13 to review the split sections.
Pre-trained weights, strict evaluation results, and download status can be found in Pre-trained Weights and Zero Training Experience . Historical checkpoints without a public download link should not be included in the run commands; use the built-in videos in the repository first, or train your own verifiable weights.
07 interactive collection and 11 visualization rollout require the available DISPLAY; long training from 08–10 requires sufficient model cache, checkpoint space, and stable power supply. All RUN_* switches are off by default. Check the commands and paths first, and then turn them on explicitly.
Re-execution and Acceptance
If the AMD teaching image does not have jupyter, nbclient, or ipykernel installed, a lightweight executor from the repository can be used. It shares namespaces in the order of Notebooks, executes ordinary Python code cells, and writes stdout/stderr back to the standard .ipynb outputs; interactive collection and long training are still determined by the switches within the Notebook.
bashpython code/execute_tutorial_notebooks.py \
notebooks/07_data_collection_and_audit.ipynb \
notebooks/08_act_training_rocm.ipynb \
notebooks/09_smolvla_training_rocm.ipynb \
notebooks/10_pi0_training_rocm.ipynb \
notebooks/11_mujoco_closed_loop_deploy.ipynb \
notebooks/12_pi0_strict_input_end_to_end.ipynb \
notebooks/13_pi05_random_position_eef_delta.ipynb \
notebooks/14_smolvla_end_to_end.ipynb \
notebooks/15_pi0_end_to_end.ipynb \
notebooks/16_act_end_to_end.ipynb
The execution result of the current submission is as follows: 07 corresponds to 11/11 code cells, 08 corresponds to 11/11, 09 corresponds to 11/11, 10 corresponds to 12/12, 11 corresponds to 10/10, 12 corresponds to 10/10, and 13 corresponds to 13/13. In the new end-to-end Notebook, 14 corresponds to 11/11, 15 corresponds to 9/9, and 16 corresponds to 9/9. The public Notebook uses variables such as $PROJECT_ROOT to represent machine-related paths, without including remote usernames, private IP addresses, or absolute local directories.