2024 Fall Robotics Colloquium: Erdem Biyik (USC)

In this talk

1. 0:55

   Human feedback can reduce robot training data

   Biyik says robots need to understand human goals across tasks such as exoskeleton control, traffic flow, and collaborative work. Because collecting robot interactions and human input is expensive, his group studies ways to use signals such as demonstrations, language, gaze, and corrections more efficiently.

2. 5:06

   Gaze guides attention during learning

   In a drawer-opening example, Biyik argues that human attention highlights task-relevant details, such as the robot’s hand and the handle. His group used 24 annotated frames to adapt a pretrained model that predicts attention, then supplied its saliency map to robot policies. The added signal improved learning efficiency in reinforcement learning and helped imitation learning across four real-robot tasks.

3. 12:50

   Comparative language gives richer feedback

   Biyik explains that demonstrations can be suboptimal because of difficult controls and human risk biases, while simple trajectory ratings can be ambiguous. His group maps trajectory differences and language descriptions into a shared latent space, allowing feedback such as requests about speed or position to guide learning. In tests, language feedback was faster than binary comparisons and users showed some preference for it, though the study included only 10 people.

4. 41:29

   One example or instruction can shape rewards

   Biyik describes using image and video-text models to reward robots for matching a goal image, a demonstrated video, or a task instruction. Video-based rewards assess the full behavior, and language instructions can support training without a robot demonstration. The approach generalized across tasks and morphologies, including examples based on a person’s video and an animation; Biyik notes that different ways of completing a task may receive similar rewards.