Less data, better robots, according to new Axis Robotics research
Robotics data startup
@axisrobotics, which counts Pi Network Ventures from
@PiCoreTeam among its seed investors, has published TUCO, a method for choosing which simulated demos to train a robot on.
Working with researchers from Stanford, Berkeley, Carnegie Mellon, and others, the team says a real robot arm trained on 10 real demos plus 600 carefully chosen simulated ones succeeded 86% of the time.
The same robot reached only about 26% when the 10 real demos were paired with a random 600, or with all 1,200 simulated demos at once. Tests covered three tasks, including stacking a cube, with 40 tries each.