Researchers led by Vedant Telrandhe have built an open, low-cost testbed to measure what it costs a robot to practise tasks without people resetting them, according to a paper on arXiv. Each attempt takes time, wears the motors and sometimes needs a person's attention, but the schedulers that pick what to practise next treat attempts as free, the authors write.

A self-built SO-101 arm runs two policies trained on demonstrations that reset each other, with a camera-based success detector and a ledger that logs time, joint travel and human help for every attempt. The team proposes choosing what to practise by the competence gained minus this cost.

In 51 minutes of robot time, a fixed retry limit did not reduce the share of time lost to failures, 36% against 27% for a feasibility rule, a gap the authors could not tell apart from chance, and saved no robot time. Concentrating demonstrations in a smaller area raised success there from 8 out of 20 to 10 out of 10.

Source: arXiv Robotics (cs.RO)

Background

Vedant Telrandhe

Robotics researcher

Robotics researcher.