Tabletop Manipulation

A robot arm fixed beside a table grasping, pushing, and arranging objects on its surface.

Tabletop manipulation is the most common experimental setup in robot manipulation research: one or two arms are fixed beside a table, cameras look at the tabletop from above, the side, or the wrist, and the task is to grasp, place, push, stack, arrange, or fold objects on the surface. It rules out mobility and navigation altogether, and the setup is easy to build and easy to reproduce, so a large number of datasets and benchmarks are built around it. CLIPort (2021), for example, uses one multi-task policy to cover 10 simulated and 9 real tabletop tasks; Google's Language-Table dataset collected nearly 600,000 language-annotated trajectories of pushing blocks around a table. The drawback is that it is fairly removed from a real home and has a small workspace, so recent research has increasingly expanded into mobile manipulation and whole-body manipulation.

ExampleSeveral colored blocks are scattered on a table and the user says, “arrange the blocks into a smiley face”; the arm pushes each block into position one at a time (a task from Language-Table).

Related

Manipulation · Pick-and-Place · Rearrangement · Mobile Manipulation · CLIPort · Language-Table

Sources

CLIPort: What and Where Pathways for Robotic Manipulation
Interactive Language: Talking to Robots in Real Time (Language-Table)