๐Ÿš€ WorldDrive Training and Evaluation

March 14, 2026 ยท View on GitHub


Stage 1: Trajectory-aware Driving World Model pretrain (Optional)

โš ๏ธ This stage is optional since a long time training.
You may skip it and directly use our released TA-DWM checkpoint.

Step 1: Download pretrained WM checkpoint

First you need to download the pretrained wm checkpoint training on nuPlan(cogvideox_2B_tadwm_stage1_pretrain)
๐Ÿ‘‰ Model

Step 2: Configure training script

After downloading, go to /data/worlddrive/navsim/scripts/training_wm/tadwm_stage.sh configure the training script.

Launch the training process:

cd /data/worlddrive/navsim/scripts/training_wm/
sh .tadwm_training.sh

Stage 2: Multi-modal Planner Training

Stage 2 consists of three steps:

  1. Cache dataset
  2. Download TA-DWM checkpoint
  3. Train planner

Step1: cache dataset(3D causal VAE latents)

You need to download the pretrained 3D Causal VAE from offical CogvideoX-2B HF
๐Ÿ‘‰ CogvideoX-2B VAE

You need to download the anchor and corresponding formated PDMS
๐Ÿ‘‰ Anchors

go to scripts/cache/run_caching_worlddrive.sh configure the training script.

Launch the multi-gpu data cache process:

sh scripts/cache/run_caching_worlddrive.sh # navtrain
sh scripts/cache/run_caching_worlddrive_eval.sh # navtest for eval

Step2: download ta-dwm checkpoint

You can download the corresponding ta-dwm checkpoint training on NAVSIM (worldtraj_stage1_1024_tadwm) or use the checkpoint from Stage 1.
๐Ÿ‘‰ TA-DWM Model

Step3: train planner

go to scripts/cache/run_worlddrive_planner.sh configure the training script.

Set:

  • checkpoint_path: TA-DWM checkpoint
  • cache_path: cached latents
  • vocab_path: anchors vocabulary

Launch the multi-gpu training process:

sh scripts/training/run_worlddrive_planner.sh

Expected behavior

During initialization, the log will show:

  • loaded weights from TA-DWM
  • newly initialized modules

Only the following module should be newly initialized:

trajplanner

Step4: evaluate planner

go to scripts/evaluation/run_worlddrive_planner_pdm_score_evaluation_stage1.sh configure the evaluation script.

Launch the multi-gpu evaluation process:

sh scripts/evaluation/run_worlddrive_planner_pdm_score_evaluation_stage1.sh