Datasets
We use the following 3 datasets for training, validation and testing which were created based on a bullet simulation. It shows a manipulation task where the gripper tries to physically separate one object from the others (singulation). We trained the gripper agent with a sparse reward function (+1 if target object has >10cm distance to all other objects, else 0) and A3C.
3 objects
3 objects
7-15 time steps
used for training & validation
5 objects (2 novel)
5 objects (2 novel)
7-50 time steps, 3 known and 2 novel shapes
used for testing only
5 objects (5 novel)
5 objects (5 novel)
7-50 time steps, 5 novel shapes
used for testing only