SenseAct™

SenseAct™ is a benchmark task suite for developing and evaluating reinforcement learning methods with physical robots by abstracting over the complexity of real-time control of robotic components. SenseAct’s guiding principles of minimizing delays and maximizing timing consistency via proactive computation lead to responsive learned behavior and reliable learning via state-of-the-art algorithms. All the task implementations share a core structure which can be reused to implement new robotic tasks that benefit from SenseAct’s tight control over system delays.

ur-reacher-6-trpo.gif
create-mover-ppo.gif
dxl-tracker-trpo.gif
 
SenseAct is an important new step in machine learning research on robots, enabling consistent and reproducible results on physical robots for the first time. I am particularly impressed by the attention that has been paid to efficient implementation minimizing the delays in the real-time learning system. I plan to adopt SenseAct in some of my own research at the University of Alberta. SenseAct will establish a standard that may greatly accelerates machine learning research on physical robots, pushing reinforcement learning to a new level of performance just as large standard data sets have for supervised learning.
— Richard S. Sutton, professor of Computing Science and AITF Chair in Reinforcement Learning and Artificial Intelligence at University of Alberta
 

AI Research Blog