Robot Learning
The course is about:
How robots learn skills
The algorithms and techniques
About the Course
Most robots today can only do simple things as programmed by professionals. If the tasks have changes, the robots need to be re-programmed. Therefore, the robots cannot adapt to product changes. With the development of Artificial Intelligence, we can teach robots to learn skills. Robotic Learning is a course covering the AI technology which is the fundamentals for the next generation robots.
Main Contents
Robots and Applications
- Manipulation
- Locomotion
- Smart manufacturing application
- Warehouse application
- Service application
Challenges with Current Deep Reinforcement Learning
- Sample complexity
- Hyperparameter tuning
- Reward specification
- Exploration
- Generalization
- Scalability
Deep Imitation Learning
- Demonstration
- DAGGER
- Few-shot imitation learning
- Policy aggregation
- Policy gradient with demonstrations
Soft Actor-Critic and Applications
- Maximum entropy RL
- Soft policy and soft actor-critic
- The optimization problem
- Soft Actor-Critic algorithm
- Applications
Meta Learning
- RL2 - Fast Reinforcement Learning Via Slow Reinforcement Learning
- A Simple Neural Attentive Meta-Learner
- Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks
Hierarchical Reinforcement Learning
- Data=efficient hierarchical RL
- FeUdal networks
Vision-based Robotic Manipulation
- Imagined goals
- QT-Opt