RI Seminar: Sidd Srinivasa: Robotic Manipulation...
Sidd Srinivasa, a robotics researcher at Carnegie Mellon’s Robotics Institute, reflects on his work in robotic manipulation and human-robot collaboration at a CMU Robotics Institute seminar.
In this talk
- 3:11
Build robots that interact with the real world
Srinivasa describes manipulation as his central research goal: enabling robots to pick up, push, and rearrange objects in physical environments. The Oreo demonstration showed how perception, motion planning, and compliant control could work together on delicate objects that vary in shape and thickness.
- 11:41
Use physics to make objects easier to grasp
Srinivasa traces his interest in physics-based manipulation from a wheeled robot that moved paper to systems that push objects into more predictable positions. His team used quasi-static pushing models to calculate capture regions, allowing robots such as HERB to funnel uncertain objects into graspable poses. The approach worked well for many tabletop objects, but depended on simplifying assumptions about physics.
- 25:06
Plan whole-arm rearrangement actions
The team extended pushing beyond the hand, using the robot’s arm and elbow to move multiple objects and clear clutter. Randomized tree-based planning searched possible robot and object configurations, while restricting motion to the table made the search more practical. The same algorithm was demonstrated on HERB and a NASA rover that used its body to reconfigure terrain.
- 41:08
Account for uncertainty and feedback
Srinivasa explains that open-loop plans can fail because they do not track uncertainty in object positions or physics. Path divergence helped identify motions that were less sensitive to small changes in starting conditions, while Monte Carlo planning offered a way to handle uncertainty more flexibly. He also describes tactile state estimation and emphasizes that feedback helps robots respond to surprises.
- 52:03
Apply manipulation to caregiving
Srinivasa discusses wheelchair-mounted robot arms that help people with high spinal cord injuries perform tasks such as retrieving items, shaving, and eating. Shared autonomy can interpret noisy user input, including brain-computer interface signals, to assist with complex actions. For feeding, the team studies food recognition, force sensing, and the social timing of when to offer a bite.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.