Assembly robots are robots that can assemble various products from individual parts. They can be very useful for industry, but for this they need to be able to manipulate objects and plan their work. So how do you test and compare the abilities of different assembler robots?

To do this, scientists from the Oxford Institute of Robotics, the Manufacturing Technology Center (MTC) and the University of Birmingham created a special benchmark , which was called RAMP (Robotic Assembly Manipulation and Planning). A benchmark is a set of tasks and criteria by which robot performance can be judged. RAMP is an open benchmark, that is, anyone can use and supplement it.

RAMP consists of a series of basic beams in different colors and sizes, which can be connected to each other. The task of the robot is to assemble a specific object from these parts, for example, a cube or a pyramid. In this case, the robot must independently choose the order of assembly of parts, so as not to get stuck in a dead end.

The scientists tested their benchmark on a real UR10 robot and in Nvidia’s Isaac Sim computer simulation. They also proposed a basic method for solving the RAMP problem based on a deep reinforcement learning algorithm. This means that the robot learns from its experience, receiving a reward for correct actions and a penalty for mistakes.

The results showed that the basic method works well in a simulation, but is not very efficient on a real robot. This suggests that the RAMP task is indeed challenging and requires further research and improvements.

“We want to create a platform that will drive the development of robotics from an applied perspective in manufacturing and assembly,” said Professor Ingmar Posner from the Oxford Institute for Robotics. He added that the RAMP benchmark is open to anyone who wants to participate in this process.