| license: mit | |
| tags: | |
| - occupancy-prediction | |
| - reinforcement-learning | |
| - autonomous-driving | |
| - stable-baselines3 | |
| - gradio | |
| library_name: voxel_car | |
| # SynthOccPred / voxel_car models | |
| This repository contains two trained artifacts for the `voxel_car` project: | |
| | Artifact | Hub path | | |
| |---|---| | |
| | Occupancy predictor checkpoint | `occupancy/ckpt_best.pt` | | |
| | PPO policy for RL planner | `rl/ppo_voxel_car_final.zip` | | |
| ## Use from Python | |
| ```python | |
| from voxel_car.hub import load_occnet_from_hf, load_ppo_from_hf | |
| model, ego_cfg, cam_cfg, image_hw, device = load_occnet_from_hf( | |
| repo_id="mmkuznecov/SynthOccPredModels", | |
| ) | |
| ppo = load_ppo_from_hf( | |
| repo_id="mmkuznecov/SynthOccPredModels", | |
| ) | |
| ``` | |
| ## Use in the Gradio demo | |
| ```bash | |
| export VOXEL_CAR_HF_REPO="mmkuznecov/SynthOccPredModels" | |
| python app.py | |
| ``` | |
| The demo can run closed-loop scenarios with either: | |
| 1. standard A* planner over OccNet occupancy predictions | |
| 2. PPO-RL planner over OccNet occupancy predictions | |