Researchers including Che Tsang Calvin, writing as the Maniformer Real-World Manipulation Team, have studied how the design of a robot data collection system affects demonstrations and the policies learned from them, according to a paper on arXiv. They combined a kinematic study of 25 dual-arm robots, teleoperation experiments and real-robot tests of fine-tuned π0.5 policies.

Controlling body parts in parallel cut task completion time by 43 to 53%, and compliant control improved success and lowered peak contact forces on tasks with tight contact. Modest shoulder motion substantially widened the useful workspace.

Demonstrations that showed the robot actively searching raised arm-selection accuracy from chance, 50%, to 100%. Miscalibration at run time made all 30 trials fail, and policies kept about 90% of the extra path length from demonstrations with detours.

Source: arXiv Robotics (cs.RO)

Background

Che Tsang Calvin

Robotics researcher

Robotics researcher.