49 min
Seminar, Hosted by Simons Institute for the Theory of Computing
Do we need diffusion in robotics?
Carnegie Mellon University
- Robotics, Machine Learning and
Recording
Abstract
Generative control policies are often credited with representing multimodal robot actions. This talk tests that explanation through experiments and theory, finding benefits instead associated with iterative computation, training noise and on-manifold biases. The results motivate alternatives that obtain those algorithmic advantages without relying on distribution matching as the explanation.
Topics
- Generative control policies, broader topic: control policies
- multimodal robot actions, search World Wide
- iterative computation, search World Wide
- training noise, broader topic: noise
- on-manifold biases, broader topic: biases
- distribution matching, search World Wide
- algorithmic advantages, search World Wide
- control policies