Robotic leg control is a standard engineering task. We’re interested in a rapid adaptation to an arbitrary leg shape/number of joints. Quality diversity algorithms offer themselves as an efficient approach to discovering control policies and robot joint configurations that achieve a diverse set of goals and target configurations, respectively. Since these algorithms are inherently easy to parallelize, we can speed up the whole process by running multiple instances of the algorithm at the same time.
The VeriDream consortium is using the videogame Space Engineers as a testing ground to run these algorithms. Space Engineers were a natural fit for the study, since its developer, Keen Software House, is GoodAI’s sister company.
You can check out an early snapshot of their progress in the video below.
The video shows large-scale parallelization of a policy search for a robotic leg. Because of the properties of the game engine, a robotic leg simulation in Space Engineers is closer to reality (and the challenges it brings) than one would think. At the same time, the game engine’s efficient code allowed for a non-trivial speedup by parallelizing the evaluations of policies offered by the policy discovery algorithm. The policy discovery relies on a quality diversity algorithm from ISIR. This algorithm has been previously successfully applied to discovering a policy of an object-throwing robot called Baxter.
The VeriDream project is an international consortium of six organizations across Europe that has been awarded €2 million by the European Innovation Council to carry out a research and innovation strategy to improve robotic performance at small and medium-sized enterprises (SMEs) using artificial intelligence (AI). The consortium is made up of:
- The German Aerospace Center (DLR)
- Sorbonne University
- Magazino GmbH
Check out the VeriDream website here.