Participatory Algorithm Design in Human-Robot Interaction

January 18, 2022
11:00am ET
Speaker: Jindan Huang
Host: Elaine Short


Quals talk:

Interactive reinforcement learning (I-RL) is a widely-used technique for robot learning. It allows the agent to receive reward signals not only from the environment but also from a human teacher. In the early stage of I-RL algorithm implementation, prior work usually tests the algorithms with simulated perfect oracles, which are assumed to know the optimal policy for any states and can provide accurate feedback to the learning agent without delay. However, a perfect oracle is not an accurate reflection of human feedback behaviors. Human teachers could respond to a learning agent in a delayed, stochastic and unreliable way. They could also give different feedback in response to the same thing because of their unique personalities, preferences and experience. To solve the above-mentioned problems, we purpose a participatory design approach that can modify perfect oracles to human-like oracles for I-RL algorithm testing. We conduct a study using our methodology. In this talk, I will present the procedure of the study and show some preliminary results. I will also discuss the improvements and future work of this research.

Please join meeting via Zoom.

Join Zoom Meeting:

Meeting ID: 971 8312 0811

Password: See colloquium email

Dial by your location: +1 646 558 8656 US (New York)

Meeting ID: 971 8312 0811

Passcode: See colloquium email