MomWorld: nuScenes-6s Checkpoint

This is the best audited MomWorld nuScenes-6s checkpoint, trained and evaluated for 6-second open-loop planning on the nuScenes benchmark. MomWorld extends MomAD with a mode-conditioned latent multi-agent world rollout and temporally gated residual planning corrections.

Checkpoint

File Training point Size SHA256
momworld_nuscenes_6s_oracle_iter_18752.pth iter 18,752 366,852,408 bytes 2c003f12a4e956eeb31f286ea01afa195b810bd859e021ae0f0dc1e2f3201376

The checkpoint was warm-started from the MomAD planner and trained with the matched oracle-mode regularized 6-second configuration included in this repository. It predicts 12 future steps at 0.5-second intervals.

Audited nuScenes validation results

Fresh evaluation used all 6,019 nuScenes validation samples.

Metric Result
Detection mAP 0.4152
NDS 0.5257
AMOTA 0.3799
Map mAP 0.5616
Planning L2 0.4945 m
Object collision 1.002%
Box collision 0.161%
Temporal consistency 0.3995

Long-horizon planning breakdown:

Horizon L2 (m) Box collision (%) Consistency (m)
4 s 1.2786 0.825 0.9271
5 s 1.7625 1.358 1.1855
6 s 2.3080 1.966 1.4593

Code and configuration

  • Code: https://github.com/modaxiansheng/MomWorld
  • Configuration: MomAD_small_stage2_MomAD_World_model_6s_v2_oracle_mode_reg02_resume_repro.py
  • Latent world model: open_loop/projects/mmdet3d_plugin/models/motion/latent_world_model_MomAD_World_model_6s.py
  • Planning head: open_loop/projects/mmdet3d_plugin/models/motion/motion_planning_head_MomAD_World_model_6s_v2_oracle_reg_backup_20260628.py

Example evaluation from the MomWorld repository root:

python open_loop/tools/test.py \
  open_loop/projects/configs/MomAD_small_stage2_MomAD_World_model_6s_v2_oracle_mode_reg02_resume_repro.py \
  /path/to/momworld_nuscenes_6s_oracle_iter_18752.pth \
  --out /tmp/momworld_results.pkl \
  --eval bbox

The nuScenes dataset, generated metadata, anchors, and evaluation outputs are not included. Follow the repository setup instructions and the nuScenes terms of use. This is a research checkpoint and should not be used as a standalone safety-critical driving system.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support