This approach may enable a robot to quickly learn to perform specific tasks in a user’s home without the owner having to physically demonstrate each task. The robot could investigate all alone, with publicly supported nonexpert input directing its investigation.
“In our strategy, the award capability directs the specialist to what it ought to investigate, rather than telling it precisely how it ought to follow through with the responsibility. According to lead author Marcel Torne ’23, a research assistant in the Improbable AI Lab, “the agent is still able to explore, which helps it learn much better” despite the fact that the human supervision is somewhat inaccurate and noisy.
Torne is joined on the paper by his MIT counsel, Agrawal; senior creator Abhishek Gupta, partner teacher at the College of Washington; as well as other individuals at MIT and the University of Washington. The study will be presented the following month at the Conference on Neural Information Processing Systems.