Facilitating Robot Learning

The ability to improve behavior through learning is the hallmark of intelligence, and thus the ultimate challenge of AI and robotics. We propose that one of the stumbling blocks in scaling up learning (to real-world problems and domains) has come from the emphasis of mathematically pure rather than practical approaches. Striving for optimality has overshadowed the plight for efficiency, and the attempt to avoid experimental bias has crippled the ability to focus on principled methods for building in domain knowledge. In what we call "great expectation", most have come to expect that a learning algorithm must achieve optimal performance with little or no built in knowledge. To make this possible, the algorithm is allowed to learn from prohibitively many examples and/or for a prohibitively long time. This trade-off between a priori bias and learning time is impractical for a great majority practical applications. Furthermore, it is counter to the way biological systems learn.

We argue that taking inspiration from biology, focusing on real-world domains and tasks, and experimentally validating all algorithms in such domains, will result in more scalable approaches. In our own work, we pursue an experimental approach in two highly uncertain, dynamic, and high-dimensional domains: multi-robot learning, and learning by imitation. Both force us to deal with perceptual and action uncertainty, non-stationarity, and real-time constraints. As a result, our approaches to learning, inspired by learning in biology, strive for efficient solutions that can cope with the challenges of real-world domains.

The key principles we take as inspiration from biological learning are: