RealAIGym: Education and Research Platform for Studying Athletic Intelligence
Felix Wiebe, Shubham Vyas, Lasse Jenning Maywald, Shivesh Kumar, Frank Kirchner
Editors: Brian Plancher, Dylan Shell, Kris Hauser, Shuran Song, Katja Mombaur
In Mind the Gap: Opportunities and Challenges in the Transition Between Research and Industry, (RSS-2022), 27.6.-01.7.2022, New York, New York, Robotics Science and Systems, Online Proceedings, 2022. RSS Foundation.
Traditional robots today (such as the ones used in factories)
have a fixed base and are fully actuated under their operat-
ing conditions. However, modern robots inspired by animals
are not bound to one place and are always underactuated.
Like animals, these robots can perform dynamic movements,
demonstrate compliance, and are robust to contact during
their movements. The interest in dynamic robot behaviors has
increased significantly due to the impressive athletic behaviors
shown by robots developed by e.g. Boston Dynamics, MIT
mini cheetah and Agility Robotics. This gives rise to
the need for canonical robotic hardware setups for studying
underactuation and comparing learning and control algorithms
for their performance and robustness. These hardware setups
and the accompanying software should be affordable, open and
accessible. Similar to OpenAIGym and Stable Baselines
which provide simulated benchmarking environments and
baselines for Reinforcement Learning (RL) algorithms, the
concept of RealAIGym (Real Athletic Intelligence Gym) is in-
troduced for benchmarking dynamic behaviors on real robots.
RealAIGym provides instructions for building reproducible
robotic systems based on Quasi-Direct Drives as well as
software to operate them for establishing a baseline for the
application of dynamic control algorithms on real hardware.
Underactuated Robotics, Reproducible Robotics