Kourosh Darvish on Shared Autonomy: From Human-Robot Collab to Humanoid Robot Teleop | AIR Seminar
Kourosh Darvish, a postdoctoral researcher at the University of Toronto and member of the Vector Institute, discusses shared autonomy, collaborative assembly, and humanoid robot teleoperation at the AI Robotics Seminar at the University of Toronto.
In this talk
- 1:46
What shared autonomy means
Darvish describes shared autonomy as a way for a person and an agent to work toward a shared objective. The robot may blend its own control with human input, while adapting to changing people, tasks, and surroundings.
- 5:33
A flexible system for collaborative assembly
Darvish presents a modular perception, planning, and action system developed during his PhD at the University of Genoa. It recognizes worker actions from smartwatch motion data, maps objects from point clouds, and uses hierarchical AND/OR graphs to represent assembly options. In a table assembly example, the system represented 255 possible ways for a person and robot to divide the work.
- 26:49
Predicting human actions and motion
Darvish describes a model that predicts both what a person will do and how they will move, using body motion and interaction data. A mixture-of-experts model combines action probabilities with motion predictions; the system ran at 25 hertz and recognized actions including walking, standing, and turning.
- 35:41
Teleoperating humanoid robots
Darvish outlines a teleoperation system that captures human movement, maps it to robot motion, and returns visual, audio, and haptic feedback. His examples include whole-body motion retargeting, walking and balance control, and a 19-degree-of-freedom hand driven by nine motors. In an ALOHA competition, the team scored 95 out of 100 in the finals.
- 53:39
Choosing autonomy for the task
In the discussion, Darvish says the right balance between teleoperation and autonomous action depends on the task and the operator’s trust in the system. Teleoperation can help in unfamiliar disaster settings, while learned action primitives may suit tasks the robot can plan and repeat. He sees humanoid form as potentially useful for varied terrain and social interaction, while noting that stability and fragility remain challenges.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.