DIVA
Released checkpoints of Beyond Planning: Differentiable Control-in-the-loop End-to-End Autonomous Driving.
Code, training and evaluation: the DIVA repository (diva/, scripts/).
DIVA keeps the planner's explicit outputs (route + target speed) and replaces the hand-tuned controller with a learned
control head that outputs throttle, brake and steering. Training rolls the head's action knots through a differentiable
vehicle model identified on the CARLA ego vehicle (vehicle.lincoln.mkz_2020, CARLA 0.9.15) and supervises the
resulting states. Planner and control head are trained together, in one stage, from scratch.
Bench2Drive-220 closed loop with diva.pt and the stop-sign rule on (single run): Driving Score 86.73, Success Rate 71.7 %.
Files
| file | size | sha256 | content |
|---|---|---|---|
diva.pt |
53 MB | 2bd65b0e0f565734f53edd7f985e1c03bcff3db38c03816695e31aa4bebc3869 |
DIVA: planning head + control head ({model, planner, cfg, ctrl_cfg}), loaded by diva.model.control_head.DIVAModel.from_checkpoint |
stop_head.pt |
19 MB | b1ff9d9e9c9a6d983fb485bfffb2fbce031f413202e7e4f4559c1d7c32101a25 |
stop-sign detector of the agent's optional stop-sign rule (DIVA_STOP_RULE=1, the default) |
calib_raw.tar.gz |
55 MB | 6370c32564f67ca768f1b5845c8ad90845a0e369bf1f66feab615b9dd41e1458 |
raw CARLA calibration logs read by the offline vehicle-identification steps (tar -xzf calib_raw.tar.gz -C data/calib) |
The TransFuser++ backbone is not included: use the all_towns weights of
carla_garage (model_0030_2.pth, the file the feature cache and the
agent use).
Usage
hf download ZhidongWang/DIVA diva.pt stop_head.pt --local-dir checkpoints
CKPT=checkpoints/diva.pt bash scripts/eval_bench2drive.sh # Bench2Drive-220, official scripts
python -m diva.tools.agent_dry_run --ckpt checkpoints/diva.pt # no CARLA server: agent driven through the vehicle model