Learning Contact-Rich Tasks with Haptic Input
Robots are typically deployed in highly controlled environments, but for robots to become usable in everyday situations, they must be able to learn to adapt to the environment. Reinforcement learning is a popular framework in robotics for teaching robots motor skills in unknown environments through a trial-and-error process. Recently, methods have been developed that allow optimization of high-dimensional control policies, which enables the use of powerful tools from machine learning such as neural networks to decide on which actions the robot should take and when. My project involves using various types of neural networks in tandem with high-dimensional sensor data (such as pressure, torque, and visual) to learn manipulation tasks.
Message to Sponsor
- Major: Electrical Engineering & Computer Sciences
- Sponsor: Rose Hills
- Mentor: Pieter Abbeel, Electrical Engineering & Computer Sciences