Talk: Autonomous robot skill acquisition in complex environments
by Dr. Herke van Hoof McGill University Montreal Canada
Robotics holds promise for lightening the workload and increasing the productivity of people performing domestic tasks or working in healthcare and small businesses.
However, due to the varied nature of tasks in such domains, automation has proven difficult and robots are hardly deployed. Deep reinforcement learning offers a framework for acquiring the required skills autonomously. However, current technologies require more data than can realistically be gathered for every new task.
In this talk, he will address how training data can be used more efficiently, and which other sources of information can be leveraged. He will discuss how skills can be learned from small sets of high-dimensional data points, by avoiding the tendency of learning systems to be overly greedy. Furthermore, he will discuss a strategy that allows robots to allocate resources more efficiently during the learning phase. As a result, skills can be acquired faster.
To scale learning methods to more complex tasks without requiring an ever-growing amount of training data, other sources of information should be used in addition to deep reinforcement learning methods. As examples, he will discuss current and planned work on using system models, simulators, or experience in related tasks to allow more efficient robot skill acquisition.
Location: UvA Science Park 904, room C1.112
Science Park 904
Science Park 904 | 1098 XH AmsterdamGo to detailpage
+31 (0)20 525 8626