Berkeley Cable Routing Multi-stage Robotic Cable Routing Task Dataset

The Berkeley Cable Routing dataset is a dataset for studying multi-stage robotic manipulation tasks, especially applied to cable routing tasks. The task requires the robot to pass the cable through a series of clamps, which represents the challenges of complex multi-stage robotic manipulation scenarios, including handling deformable objects, closing visual perception loops, and handling extended behaviors consisting of multiple steps. The dataset was released in 2023 by a research team from the University of California, Berkeley and Intrinsic Innovation LLC. The related paper results are "Multistage Cable Routing Through Hierarchical Imitation Learning".

In this dataset, the researchers present an imitation learning system that uses vision-based policies that are trained with demonstrations from lower (motor control) and higher (sequencing) levels. The system is able to recover from failures and compensate for the deficiencies of the low-level controllers by intelligently choosing which controller to trigger, retrying, or taking corrective actions. The researchers show that the system performs well in generalizing to challenging variations in gripper placement.

This dataset opens up 3 sets of data in the project:

• Routing Primitives Offline Dataset: 1,442 low-level routing traces that can be used to train routing policies.

• High-Level Primitive Selection Offline Dataset: 11,915 transformations and their associated observations used for training high-level primitive selection policies in research teams’ systems.

• End-to-End Trajectory Dataset: 257 trajectories demonstrating an end-to-end task, routing the cables through all the clips on the board.