Many complex skills and behaviors are difficult to express programmatically. Reinforcement learning is a trial-and-error system for learning optimal behaviors where a behavior is optimized through exploratory interactions with an environment. Hence, reinforcement learning can provide an agent with the capability to practice and learn a skill autonomously. However, when applied to physical systems such as robots, reinforcement learning can be costly due to long learning times and unsafe due to the required exploration.
What we do
- Learning in-contact skills, where we have developed reinforcement learning for complex in-contact skills such as wood planing;
- Learning grasping and reaching behaviors, where we have developed methods for safe exploration;
- Incremental learning, where we have developed learning of incrementally complex skills;
- Skill generalization, where we have developed methods to generalize skills to new situations.
Getting robots to autonomously learn to perform various tasks is often a long-term process, during which the robot’s exploratory actions can be unpredictable and potentially dangerous to the surrounding environment and to the robot itself. To mitigate the risk of hardware damage and to speed up the learning process, initial phases of learning are often […]