README

1. Setup

This project uses uv for dependency management with a workspace structure.

# Sync workspace and create virtual environment
uv sync

# Activate the virtual environment
source .venv/bin/activate

# Install git hooks
uv run pre-commit install

# check torch
python3 -c "import torch; print(torch.cuda.is_available())"

Run all configured hooks manually with:

uv run pre-commit run --all-files

Autoresearch Control Panel

The autoresearch workflows can be launched from a unified Gradio control panel:

source .venv/bin/activate
python -m control_panel

See control_panel/README.md for the workspace layout, asset registry, training/eval tabs, PRiSM flow, Perception Reproducer route mining, rendering, and Scene Editor integration.

2. Create dataset

2.1. Prepare rosbags

We assume the following directory structure:

driving_dataset$ tree . -L 2
.
├── bag
│   ├── 2024-07-18
│   │ ├── 10-05-28
│   │ ├── 10-05-51
│   │ ├── ...
│   │ ├── 16-10-07
│   │ └── 16-27-15
│   ├── 2024-12-11
│   ├── 2025-01-24
│   ├── 2025-02-04
│   ├── 2025-03-25
│   └── 2025-04-16
└── map
     ├── 2024-07-18
     │   ├── lanelet2_map.osm
     │   ├── pointcloud_map_metadata.yaml
     │   ├── pointcloud_map.pcd
     │   └── stop_points.csv
     ├── 2024-12-11
     ├── 2025-01-24
     ├── 2025-02-04
     ├── 2025-03-25
     └── 2025-04-16

2.2. Convert to diffusion_planner's format (npz)

use parse_rosbag_for_directory.py directly.

python3 ./ros_scripts/parse_rosbag_for_directory.py <target_dir_list> --save_root <save_root> [--step <step>] [--limit <limit>]

2.3. Generate path_list.json

This script search *.npz files and create path_list.json.

python3 ./diffusion_planner/util_scripts/create_train_set_path.py <root_dir_list>

3. Train

Run (launches train_predictor.py across all visible GPUs):

cd ./diffusion_planner
python3 train_run.py \
    --exp_name <exp_name> \
    --train_set_list <train.json> \
    --valid_set_list <valid.json>
    # optional: --resume_model_path <.pth> --wandb_run_id <id> --wandb_project_name <name>