Scene-aware Generative State Abstractions Dataset

Year: 2024
DOI: 10.3929/ethz-b-000634113

This dataset accompanies the publication On Learning Scene-aware Generative State Abstractions for Task-level Mobile Manipulation Planning (IROS 2024). It provides paired demonstrations and perception features for two predicates:

• on_clutter: 20,000 training and 2,000 test samples
• inside_drawer: 20,000 training and 2,000 test samples

Each sample contains:

1. Demonstration data from the PyBullet simulator, including full physics state and scene object metadata.
2. Bounding-box features extracted as described in the paper.
3. Individual object point clouds for training point-cloud auto-encoders.
4. Scene-level point clouds containing all objects for scene encoding.
5. Encoded point clouds generated with multiple point-cloud encoders.

The accompanying code for handling and experimenting with the data is available at https://github.com/ethz-asl/predicate_learning.

Data Access

➡️ ETH Research Collection / DOI landing page: https://doi.org/10.3929/ethz-b-000634113

Citation

Please cite the dataset and paper as:

J. F\u00f6rster, J. J. Chung, L. Ott, R. Siegwart, "On Learning Scene-aware Generative State Abstractions for Task-level Mobile Manipulation Planning", IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2024, pp. 13599\u201113606. DOI: https://doi.org/10.3929/ethz-b-000634113.

Tags: manipulation, planning, simulation, point-clouds, predicates, mobile-robotics