Skip to content

DATA.md

Latest commit

 

History

History
136 lines (88 loc) · 3.91 KB

DATA.md

File metadata and controls

136 lines (88 loc) · 3.91 KB

Data and Model Assets

This document explains which large files are excluded from the repository, where they can be obtained, and how to regenerate them from scratch.

What is excluded

Directory / File Size Reason
dataset/ ~420 MB Large training datasets (NPZ format)
training_data/ ~150 MB BC checkpoint grid (30+ hyperparameter runs)
trained_models/ ~150 MB RL policy checkpoints
results_pil/ ~206 MB PIL hardware simulation outputs
results_sil/ ~21 MB SIL software simulation outputs
montecarlo_results/ ~17 MB Monte Carlo evaluation results
hp_study/ ~33 MB Hyperparameter study outputs
weight_reward_study/ ~17 MB Reward weight tuning outputs
pil_simulation/env_pil/ ~180 MB ARM64 Python environment for Jetson
pil_simulation/env_onnx_7w/ ~46 MB ARM64 ONNX Runtime environment for Jetson
onnx_models/ — ONNX policy model (regenerable)
pil_simulation/models_trt/ — TensorRT engines (platform-specific)

Small files included in the repository

The following reference data files are small enough to be committed directly:

File Size Description
nominal_trajectory/mars_nominal_trajectory.npz ~700 KB Nominal Earth–Mars reference trajectory
nominal_trajectory/mars_nominal_trajectory_fixed_time.npz ~24 KB Fixed-time variant
earth_mars_nominal.npz ~41 KB Earth–Mars nominal state pair
kernels/de432s.bsp ~9 MB JPL DE432 solar system ephemeris
kernels/naif0012.tls ~5 KB NAIF leap-second kernel
kernels/pck00010.tpc ~119 KB Planetary constants kernel

Regenerating the full pipeline

Follow these steps to reproduce all large files from scratch.

Step 0 — Install dependencies

conda install -c conda-forge heyoka "heyoka.py>=5.0"
pip install -r requirements.txt

Step 1 — Nominal trajectory

Run notebooks/data_generation/Mars_nominal_trajectory.ipynb.

Outputs: nominal_trajectory/mars_nominal_trajectory.npz

Step 2 — Training dataset

Run notebooks/data_generation/Compute_Database_Mars_New.ipynb.

Outputs: dataset/mars_nominal_trajectory_*_dataset.npz (~420 MB total)

Step 3 — Behavioural cloning

Run notebooks/bc/Pretrain_Actor.ipynb.

Configure the hyperparameters at the top of the notebook. Outputs: training_data/<model_name>/ckpt/actor_best.pth

Step 4 — PPO fine-tuning

Run notebooks/rl/Train_Model_RL.ipynb.

Point to the BC checkpoint from Step 3. Outputs: trained_models/<run_name>/

Step 5 — ONNX export

Run notebooks/deployment/Onnx_export.ipynb.

Outputs: onnx_models/policy_OAM.onnx

Step 6 — SIL evaluation

Run notebooks/evaluation/SIL_simulation.ipynb.

Outputs: results_sil/

Step 7 — Monte Carlo evaluation

Run notebooks/evaluation/Montecarlo_Policy_Evaluation.ipynb.

Outputs: montecarlo_results/

Step 8 — PIL evaluation (requires Jetson hardware)

  1. Transfer onnx_models/policy_OAM.onnx to the Jetson.
  2. On the Jetson, start the inference server: 
    python pil_simulation/server_jetson_inference_onnxruntime_7W_final.py
  3. On the host, run notebooks/evaluation/PIL_simulation.ipynb.

Outputs: results_pil/

Final trained model

The production model used for PIL experiments is:

onnx_models/policy_OAM.onnx

Architecture: MLP 64×64×64, tanh activations, 8-dimensional input, 3-dimensional output.

See notebooks/deployment/Onnx_export.ipynb for the export configuration and MODEL_CARD.md (in onnx_models/) for full model details.

Git LFS

If you are storing large files in this repository via Git LFS, run the following after cloning:

git lfs install
git lfs pull

The .gitattributes file already configures which file extensions are tracked by LFS (.npz, .pth, .onnx, .engine, .bsp, .pdf).