MIT Engineers Devise a Method to Enhance Autonomous Robot Performance

Researchers from MIT have developed an optimization tool that improves the performance of virtually any autonomous robot during its design phase. Engineers often run countless simulations tailored to specific robotic processes during their design phase to achieve the best possible performance when carrying out tasks. While those ad-hoc processes are designed to bring about increased efficiency, they can’t be tailored to other robots designed to carry out different tasks. The team overcame that issue by designing a multipurpose tool to increase performance no matter the autonomous platform.

The key to efficiency lies within an optimized framework that can automatically find and tweak autonomous systems to achieve the desired outcome. The foundation of that code is based on automatic differentiation, or “autodiff,” a programming tool developed within the machine learning community and initially used to train neural networks. Autodiff is a technique that can quickly and efficiently “evaluate the derivative,” or the sensitivity to change of any parameter in a computer program.

“Our method automatically tells us how to take small steps from an initial design toward a design that achieves our goals,” explained MIT grad student Charles Dawson. “We use autodiff to essentially dig into the code that defines a simulator and figure out how to do this inversion automatically.” The engineers tested their tool using two different autonomous systems. The first was comprised of a wheeled robot tasked with planning a path between two obstacles based on signals transmitted from two positions. The team wanted to find the optimal placement for those signal beacons, which the toll optimized in about 5 minutes, compared to the 15 it would typically take.

The second system was more complex and involved a pair of wheeled robots working together to push a box to a targeted position. A simulation of this system showed many more parameters and subsystems compared to the previous experiment. Regardless, the tool identified the steps needed for the robots to accomplish their goal, which was 20 times faster than other optimization techniques. The researchers have made the optimization tool available for download on GitHub for anyone interested.