Control System From MIT Supports a New Kind of Robot

Researchers at MIT are developing a new control system to apply the agility of cheetahs to four-legged robots.

“It is remarkable to witness the flexibility of machine learning techniques capable of bypassing carefully designed intermediate processes (e.g. state estimation and trajectory planning) that centuries-old model-based techniques have relied on," says Sangbae Kim  , professor of mechanical engineering.

Transcript

00:00:01 NARRATOR: A new control system designed by a team of researchers at MIT's Improbably AI Lab enables four-legged robots to traverse across discontinuous terrain in real time. They tested their system on the MIT mini cheetah, a powerful, agile robot built by their collaborators at the Biomimetic Robotics Lab at MIT. Unlike other methods for controlling a quadruped robot, this new system does not require the terrain

00:00:26 to be mapped in advance. They're novel control system is split into two parts, one that processes real time input from the video camera mounted on the front of the robot and another that translates that information into instructions for how the robot should move its body. A novelty of their new system is that it adjusts the robot's gait. If a human were trying to leap across a wide gap, for example, they might start by running to build up speed and have

00:00:52 a powerful leap across. In the same way, the robot can adjust the timings and duration of its foot contacts to better traverse the terrain. The way in which the two separate parts or controllers work together make this system especially innovative. There are robust and effective blind systems where the robot isn't using vision, however, they only enable robots to walk over continuous terrain and other systems that do incorporate vision usually rely on a height map of the terrain

00:01:18 which must either be preconstructed or generated on the fly. A process that is slow and prone to failure if the height map is incorrect. Essentially, this new system combines the best elements from blind systems with a separate module that handles vision in real time. They conducted simulations of the robot running across hundreds of different discontinuous terrains. And, over time, their algorithm learned

00:01:40 which actions were successful. For example, when the robot approaches a gap, it can compute, based on learned behavior, the best way to move across using visual information as its guide. While they were able to demonstrate that their control system works in the lab, the researchers note it still has a long way to go before it can be deployed in the real world. In the future, there are plans to increase the robot's computing power and to improve its ability to navigate

00:02:04 in different lighting conditions, thus increasing its ability to maneuver in real time. [MUSIC PLAYING]