December 5, 2022


Sapiens Digital

MIT Engineers Showed a Robotic Cheetah How to Take One Giant Leap

The Massachusetts Institute of Technology (MIT) engineers have developed a ‘mini cheetah’ that can jump over irregular gaps with high speed and agility — and it can do it all in real-time!

The engineers installed a new technology that relies on a real-time video sensor that detects potential obstacles like gaps and holes and interprets them into instructions on how the cheetah should act to make this possible, according to an MIT press release. This vision system focuses on the depth of incoming terrain, which is sent into a neural network that “learns” from the experience and sends a target trajectory to a low-level controller that handles the bot’s 12 joints based on physical equations that describe the motion.

The innovative two-part technology means that the robot can go anywhere without needing to map the terrain first. And in the future, this could allow robots to venture into the woods on an emergency response mission or climb stairs to deliver medication to those in need.

An algorithm that ‘sees’

The researchers merged the best aspects of these robust, blind controllers with a separate module that handles vision in real-time to create their system.

“The hierarchy, including the use of this low-level controller, enables us to constrain the robot’s behavior so it is more well-behaved. With this low-level controller, we are using well-specified models that we can impose constraints on, which isn’t usually possible in a learning-based network,” explains Gabriel Margolis, a Ph.D. student in the lab of Pulkit Agrawal, professor in the Computer Science and Artificial Intelligence Laboratory at MIT, who took part in the study.

To train the neural network (high-level controller), the researchers utilized a trial-and-error method known as reinforcement learning by running simulations of the robot running across hundreds of different irregular terrains and rewarding it for successful crossings. This enabled the algorithm to learn which actions to do to get the most reward.

Then, the experiment was carried to the real world by using a set of wooden planks. They made a physical terrain and tested their control system with the mini cheetah, and it successfully traversed 90 percent of the terrains.


This comes hot on the heels of the news that researchers have attached a gun to a robotic dog just a week ago. And with robotics technologies advancing at this rate, it’ll probably only be a matter of time before the cheetah leaps out of the lab.

Source Article