Robot Highlight—Mini Cheetah Hits New Speeds
Highlighting recent robotics research, MIT's robot "mini cheetah" combines the best of electronics and machine learning to zoom towards the future.
Four-legged animals have long been a popular platform for basing walking robots on. Some of the most widely internet-famous robots are the quadrupeds that have come out of the Massachusetts Institute of Technology (MIT), such as Boston Dynamics' Spot (a spin-off of MIT bought by Hyundai) and the MIT Mini Cheetah.
MIT's mini cheetah. Image used courtesy of MIT
Despite the interest and research into quadruped robots, one of the biggest problems these robots face is the limits in their control systems when it comes to unknown terrains. Typically, the algorithms they use require continuous terrain or rely on a pre-generated heightmap of a terrain. Some systems can generate these heightmaps on the spot; however, this process is slow and prone to error since incorporating vision into a robot for traversing new discontinuous terrain is quite a difficult and intensive task.
Hoping to progress the mobility of four-legged robots, a team of MIT engineers has developed a new type of control system that has the potential to give them concurrent terrain data generation and traversing abilities.
Mini Cheetah's Autonomous Control System
Ph.D. student Gabriel Margolis and professor Pulkit Agrawal have a solution for bringing a more precise vision to quadrupedal robots. They've developed a control system, which they are calling "Depth-based Impulse Control" (DIC) containing two parts, one that processes real-time data and one that translates that data into commands, capable of improving the speed and agility while decreasing errors in legged robots.
A general joint trajectory generator (left) vs MIT's DIC system. Image used courtesy of Margolis et al
This system was tested on MIT's Mini Cheetah using depth cameras to generate concurrent data without relying on a heightmap or taking time to process the whole terrain. The controller software itself is a trained neural network (NN) that learns from its previous mistakes in a trial-and-error-based manner.
Despite the steps that this new system could have for future endeavors, this system still faces the challenge of real-world sensor interference, which can't really be factored into computer simulations.
One of the problems is that the robot's state estimator isn't powerful enough at this point to give the true positions of the Mini Cheetah. Due to this, for some of their experiments, the MIT team used external motion capture to gather high-precision and more reliable position data, which improved the system's overall accuracy.
Using this novel controller, the Mini Cheetah adjusted its gait and traversed 90 percent of the terrains simulated by the MIT team, which the engineers built out of wooden planks with gaps in between them.
Mini Cheetah leaping. Image used courtesy of Margolis et al
According to Margolis, while the system works in a lab setting, the underlying challenges need to be addressed before any real-world use. Their plan for the future is to incorporate a more robust onboard computer into the system to ditch any external motion capture arrangements.
The Future of Robotic Control Systems
While television shows and movies might often make bio-inspired robots seem scary and dangerous, the fact of the matter is that this engineering discipline not only has the potential to improve our lives but also expand and further our knowledge of both the world around us as well as emerging technologies in automation and AI.
A system such as this one could be a step toward real autonomous walking robots that can gauge their environment and decide how to perform their next step in real-time.
While the MIT team still has a lot of work to do to make their mini "cheetah" as capable as a real cheetah, their control system has the potential to revolutionize multiple types of bio-inspired robots and give them vastly improved locomotion skills for universally unexplored terrains.