The workshop Active Learning in Robotics: Exploration Strategies in Complex Environments was held at Humanoids conference Madrid on November 18, 2014.
This workshop was supported by the German Research Foundation through the Priority Programme Autonomous Learning. More information on the workshop’s website.
Robotics has achieved tremendous results: Robots can drive cars, precisely build goods at industrial scale, and perform surgery. The design of agents that can learn such complex skills by themselves is still in its infancy. One bottleneck is the lack of labeled data. Datasets are often limited in size and coverage, since they need experiments on robots or expensive simulations. Thus an important capability of every robot in complex environments is to gather new data.
To do this efficiently, intelligent robots need to be able to identify which task-relevant information is still missing. Only this ability enables agents to actively work towards a better understanding of the environment, the agent’s state, and the task at hand, which will eventually lead to better performance.
Active learning is an approach to gain this ability. Within this paradigm the agent chooses the next datapoint to achieve the best learning result with as few data points as possible. Such strategies can shorten the process of data gathering significantly and also lead to the important information more quickly. In robotics, active learning approaches can be applied on various levels, from low-level tasks such as motor control learning to higher level reasoning tasks.
In this workshop, we want to discuss the state of the art of active learning in robotics, but also address important questions that are still open. The topics discussed in this workshop include (but are not limited to):
What representation of knowledge allows efficient reasoning about it?
What are successful strategies based on these representations?
How can we generalize or transfer experiences to decrease the need of data?
How to explore safely without damaging the robot and its environment?
How can existing approaches be `scaled up’ to large, continuous real world domains?