UIUC Scooping Dataset (Granular Materials Manipulation Dataset with Scooping/Digging/Excavation Action)

Overview:
This dataset encompasses a compilation of 6,700 executed scoops (excavations), mapped across a vast spectrum of materials, terrain topography, and compositions.

Motivation:
The primary motivation behind collecting this dataset is to advance research in granular material manipulation and meta-learning. We also believe that this dataset would help with research in domain specific problems including autonomous terrain excavation. The diverse nature of the dataset, spanning a broad spectrum of materials and terrains, provides a rich foundation for various studies.

Dataset Characteristics:

• Total Samples: 6,700

• Tasks: 67 (with each task being defined by its unique combination of materials and composition)

• Materials: 12 distinct materials

• Terrain Compositions: Materials are combined in 4 different ways to create diverse terrains.

• Volume Metrics:

• Maximum scoop volume: 260.8 cm³

• Average scoop volume: 31.3 cm³

Content Summary:

For each scoop action in the dataset, the following data points are recorded:

• Action Information: This includes the scoop location, scoop yaw, scoop depth, and scoop stiffness.

• Terrain State (Pre-Scooping): Captured as RGBD data, using an overhead RealSense L515.

• Volume Scooped: Represents the volume of the material scooped during a particular action.

• End-Effector Sensor Data: F/T sensor data recorded while executing the scooping action. The scoops are executed using the UR5e.

Potential Use Cases:

1. Granular Material Manipulation: Given the varied materials and terrains, this dataset is a valuable asset for those looking into the dynamics and intricacies of manipulating granular substances.

2. Meta-learning: With 67 tasks based on materials and compositions, researchers can delve into meta-learning applications, especially regarding how different materials interact and behave in a few-shot setting.

3. Self-supervised Learning: The extensive number of samples can be harnessed for training models in a self-supervised manner, extracting patterns without explicit labeled supervision.

4. Visual Estimation of Terrain Properties: The RGBD data offers an opportunity for training models that can visually estimate and interpret terrain characteristics based on depth and color data with the F/T sensor data acting as ground truth.

5. Autonomous Excavation and Construction Robotics. The dataset provides results of scooping (excavation) on terrains with different materials including sand, rigid rocks, rock fragments, etc. and can be useful for developing novel approaches for autonomous excavation and construction robotics in general.

6. Deep Reinforcement Learning and Decision Making. The dataset provides a set of observations, actions, and the corresponding reward values (the scoop volume) and can be naturally posed as a sequential decision making problem.