AdaOcc checkpoints

AdaOcc: Adaptive 3D Occupancy Prediction for Embodied Tasks

🎉 Accepted to NeurIPS 2026.

• Project page (videos and real-robot demos): https://wangjl-nb.github.io/AdaOcc_web/
• Code, data preparation, and reproduction docs: https://github.com/wangjl-nb/AdaOcc

This Hugging Face repository only hosts the public AdaOcc checkpoint assets. It does not include OccScanNet data, generated labels/depth maps, RADIO weights, the Depth-Anything-V2 fine-tuned checkpoint, or upstream MapAnything assets referenced by the config snapshots.

Uploaded files

file	description	target path in the AdaOcc repo
pretrain/fusion_pretrain_model.pth	Slim fusion pretrain initializer for training AdaOcc from scratch.	pretrain/fusion_pretrain_model.pth
checkpoints/adaocc_online_depth_occscannet_mini_epoch200.pth	Trained AdaOcc online-depth OccScanNet-mini epoch-200 checkpoint for direct evaluation.	checkpoints/adaocc_online_depth_occscannet_mini_epoch200.pth
checkpoints/adaocc_radio_occscannet_full_epoch100.pth	Trained AdaOcc RADIO OccScanNet full-split epoch-100 checkpoint for direct evaluation.	checkpoints/adaocc_radio_occscannet_full_epoch100.pth
configs/radio_occscannet_mini_training_snapshot.py	Config snapshot from the released training run.	reference only
configs/radio_occscannet_full_training_snapshot.py	Config snapshot from the released full-split training run.	reference only
logs/online_depth_occscannet_mini_epoch200.log	Training/evaluation log for the released checkpoint.	reference only
logs/radio_occscannet_full_epoch100.log	Training/evaluation log for the released full-split checkpoint.	reference only
SHA256SUMS	Checksums for hosted assets.	reference only

Released checkpoints

dataset split	checkpoint	epoch	mIoU	IoU
OccScanNet-mini validation	checkpoints/adaocc_online_depth_occscannet_mini_epoch200.pth	200	58.49	65.49
OccScanNet full validation	checkpoints/adaocc_radio_occscannet_full_epoch100.pth	100	59.67	65.29

OccScanNet full-split checkpoint

The full-split checkpoint, config snapshot, and training/evaluation log come from the same run. The epoch-100 validation line reports mIoU=59.67 and IoU=65.29; this is also the best validation result recorded in that log.

Here, full refers to the full OccScanNet data split, not to an ablation label. The run uses RADIO features, containment loss, and a progressive query schedule from 200 to 500 queries. The released config is a snapshot with machine-local paths and is provided for reference and reproducibility auditing.

The GitHub project now ships a matching evaluated config at configs/occscannet/radio_occscannet_full.py. It is the full-split companion of configs/occscannet/radio_occscannet_mini.py with the same model, the full PKLs, the 200 -> 500 query schedule, and generated precomputed depth enabled by default.

Evaluate a released checkpoint

Run from the AdaOcc GitHub repository root after downloading the assets. Both released checkpoints load into their public configs without key remapping.

OccScanNet-mini (online Depth-Anything depth):

ADAOCC_ONLINE_DEPTH=1 \
./dist_val.sh 8 configs/occscannet/radio_occscannet_mini.py \
  checkpoints/adaocc_online_depth_occscannet_mini_epoch200.pth

OccScanNet full validation (generated precomputed depth is the config default):

./dist_val.sh 8 configs/occscannet/radio_occscannet_full.py \
  checkpoints/adaocc_radio_occscannet_full_epoch100.pth

Generate the full-split PKLs and check the checkpoint/config pair with:

python scripts/generate_occscannet_mini_pkls.py \
  --data-root data/OccScanNet \
  --train-count 0 --val-count 0 \
  --train-output train_occscannet_full.pkl \
  --val-output val_occscannet_full.pkl \
  --test-output test_occscannet_full.pkl \
  --overwrite

python scripts/check_checkpoint.py \
  --config configs/occscannet/radio_occscannet_full.py \
  --checkpoint checkpoints/adaocc_radio_occscannet_full_epoch100.pth

See docs/REPRODUCIBILITY.md, docs/DATA.md, and docs/AI_REPRODUCTION.md in the GitHub project for the complete recipe.

Download

Run from the AdaOcc GitHub repository root to download the OccScanNet-mini release:

hf download wjldragon/AdaOcc \
  pretrain/fusion_pretrain_model.pth \
  checkpoints/adaocc_online_depth_occscannet_mini_epoch200.pth \
  --local-dir .

To download the OccScanNet full-split release together with its config snapshot and log:

hf download wjldragon/AdaOcc \
  checkpoints/adaocc_radio_occscannet_full_epoch100.pth \
  configs/radio_occscannet_full_training_snapshot.py \
  logs/radio_occscannet_full_epoch100.log \
  --local-dir .

Use --local-dir . so the checkpoint paths are restored exactly under pretrain/ and checkpoints/. Do not use --local-dir pretrain or --local-dir checkpoints, which would create nested paths such as pretrain/pretrain/....

Fusion pretrain note

pretrain/fusion_pretrain_model.pth is a slim OPUS-derived initializer. It keeps the sparse middle-encoder weights used by AdaOcc's public config (pts_middle_encoder.*) and removes unused branches. The extraction script is in the GitHub project at scripts/extract_adaocc_fusion_pretrain.py.

For data preparation, environment setup, training, evaluation, and upstream asset instructions, please use the GitHub project: https://github.com/wangjl-nb/AdaOcc.

Citation

If you find AdaOcc useful in your research, please consider citing our paper:

@inproceedings{wang2026adaocc,
  title     = {AdaOcc: Adaptive 3D Occupancy Prediction for Embodied Tasks},
  author    = {Wang, Jinglong and Wang, Yunjie and Zhang, Zhiyang and
               He, Jiawei and Yuan, Ye and Qiu, Bo and Zhang, Jing},
  booktitle = {Advances in Neural Information Processing Systems},
  volume    = {39},
  year      = {2026},
  note      = {Accepted to NeurIPS 2026}
}

This entry is valid now and will be updated with the official proceedings key, pages, and URL once the NeurIPS 2026 proceedings are published.