Robotic Scooping Data for Granular Material Manipulation

6,700 samples collected over 67 distinct terrains for robotic manipulation of granular materials. The dataset, created by pthangeda, includes terrain metadata, RGB images, and 16-bit depth images of terrains before scooping. It was last updated on October 4, 2024.

Use Cases

• Train a model to predict scooping success from Terrain ID, composition, and materials using the terrain metadata.
• Develop computer vision models for terrain analysis by processing the provided RGB Image and Depth Image pairs.
• Benchmark reinforcement learning algorithms for robotic arm control using the 6,700 recorded scooping attempts.
• Study the relationship between terrain properties (composition, materials) and the required scooping action from the associated sensor data.

Strengths

• Contains 6,700 individual samples for model training.
• Covers 67 different terrain configurations for evaluating generalization.
• Provides aligned RGB and 16-bit depth image data for each sample.

Limitations

• Sample size of 6,700 may be limited for training large, data-hungry neural networks.
• Lacks explicit information on scooping outcomes or action labels, limiting supervised learning approaches.
• Geographic origin and specific robotic hardware used for data collection are not specified.

Provenance

Source

pthangeda on Hugging Face.

Collection Method

Collected via robotic arm experiments on constructed terrains.

Freshness

Last updated in October 2024.