See2Act demonstrations

Scripted keyframe demonstrations for the four occluded Ravens tasks of Learning to See While Learning to Act: Diffusion Models for Active Perception in Robot Imitation (Kuancheng Wang, Vaibhav Saxena, Shuo Cheng, Yotto Koga, Danfei Xu; arXiv:2606.23625). Collected with scripts/collect_demos.py of github.com/KuanchengWang/see2act; the trained policies are at harrywang01/See2Act.

file	task	demonstrations
place-red-in-green.hdf5	red block and green bowl among eight distractors	100 train + 20 validation
bin-picking.hdf5	the block sits inside a bin lying on its side	100 train + 20 validation
put-within-shelf.hdf5	block on the top level, bowl on the bottom level of a shelf	100 train + 20 validation
bin-search.hdf5	three bins, the block is inside one of them	100 train + 20 validation

Training demonstrations use even seeds and validation demonstrations odd seeds; only successful demonstrations are kept and the seed of each one is stored with it.

Format

Each file is an hdf5 dataset with one group per demonstration:

data.attrs["task"]                         task name
data/demo_i.attrs["seed"]                  episode seed
data/demo_i/actions          (1, 14)       pick position (3), pick quaternion (4), place position (3), place quaternion (4)
data/demo_i/scene            (8 k,)        all object poses of the scene (name, position, quaternion), see see2act/sim/scene.py
data/demo_i/obs/rgb_top      (1, 320, 320, 3) uint8   reference image, top-down camera
data/demo_i/obs/rgb_front    (1, 320, 320, 3) uint8   reference image, front camera
mask/train, mask/valid                     demonstration names of each split

The policy never reads the stored images: during training and inference the renderer rebuilds the scene in PyBullet from scene and renders the views prescribed by the camera schedule.

Usage

git clone https://github.com/KuanchengWang/see2act.git && cd see2act   # install per the README
hf download harrywang01/See2Act --repo-type dataset --local-dir data
python scripts/train.py --config configs/see2act.json --dataset data/bin-picking.hdf5 --name see2act_bin-picking

Integrity

b4d916e61f48c550cef0ab7fc1ebca97  bin-picking.hdf5
4498b551a7e1c877f4c17ffd25a9d48e  bin-search.hdf5
3dd6c442fcd470c518007a390aacc168  place-red-in-green.hdf5
d9cd37f8fc595ecb67c225609f514adb  put-within-shelf.hdf5

Citation

@misc{wang2026learninglearningactdiffusion,
      title={Learning to See While Learning to Act: Diffusion Models for Active Perception in Robot Imitation},
      author={Kuancheng Wang and Vaibhav Saxena and Shuo Cheng and Yotto Koga and Danfei Xu},
      year={2026},
      eprint={2606.23625},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2606.23625},
}