A new control method, demonstrated with MIT’s robotic mini cheetah robot, allows four-legged robots, using their legs, to leap across uneven terrain in real-time.
A bounding cheetah runs across a rolling field, avoiding gaps in the rough terrain. Although it may seem effortless, getting a robot moving in this manner is different.
Four-legged robots inspired by the movements of cheetahs or other animals have made considerable strides in recent years. Still, they trail their mammalian counterparts when they traverse landscapes with rapid elevation changes.
You must use your vision to avoid failure in those situations. If you have an idea, it is easier to avoid falling into a gap. Some methods can be used to incorporate the concept into leg locomotion. Still, most are not suitable for emerging agile robotic systems,” said Gabriel Margolis, a Ph.D. candidate in the lab Pulkit Agrawal at MIT. Margolis is also a professor in the Computer Science and Artificial Intelligence Laboratory CSAIL.
Margolis and his colleagues have now developed a system to improve the speed and agility of legged robots when they jump across obstacles in the terrain. The control system’s innovative design is divided into two parts. One processes real-time input from the robot’s front camera, and the other converts that information into instructions on how to move the robot’s body. Researchers tested the system on the MIT Mini Cheetah robot, a powerful and agile robot made in the laboratory of Sangbae K, professor of mechanical engineering.
This two-part system is unlike other methods of controlling a four-legged robot. It doesn’t require that the terrain be planned. The robot can move anywhere. This could allow robots to run into the woods for an emergency response mission or climb stairs to deliver medication to elderly residents.
Margolis co-authored the paper with Pulkit Agrawal (senior author), who is the Improbable AI Lab at MIT, and Steven G. and Renee Finn, career development assistant professors in the Department of Electrical Engineering and Computer Science. He was also Professor Sangbae Kim at MIT and fellow graduate students Tao Chen and Xiang fu at MIT. Kartik Paigwar (a graduate student at Arizona State University) and Donghyun K (an assistant professor at the University of Massachusetts at Amherst) are other co-authors. Next month, the work will be presented at the Conference on Robot Learning.
All is under control.
This system is particularly innovative because it uses two controllers that work together.
A controller is an algorithm that will transform the robot’s current state into a series of actions it can follow. Blind controllers, which do not include vision, are durable and efficient but allow robots to walk on continuous terrain.
These algorithms cannot handle vision because it is a complicated sensory input. Most systems incorporating ideas rely on a “heightmap,” a terrain map. This must either be preconstructed or created on the spot. If the height map needs to be corrected, it can lead to slow processing and failure.
The researchers combined the best parts of these blind controllers with a module that can handle vision in real time to create their system.
The robot’s camera takes depth images of the terrain ahead and feeds them to a high-level controller. Also, information about the robot’s state (joint angles and body orientation) is provided. A neural network is a high-level controller. It “learns from” experience.
The neural network produces a target trajectory the second controller uses to calculate the torques for each robot’s 12 joints. This controller at the low level is not connected to a neural network but instead uses a series of simple physical equations to describe the robot’s motion.
“The robot’s behavior can be controlled by the hierarchy, which includes the use of the low-level controller. Margolis states that the low-level controller allows us to use well-defined models to place constraints. This is not possible in a learning-based system.
Teach the network
To train the controller’s high-level, the researchers used reinforcement learning (a trial-and-error) to teach it. The researchers ran simulations that showed the robot traversing hundreds of discontinuous terrains. They rewarded the robot for crossing them successfully.
The algorithm learned over time which actions yielded the greatest reward.
They then created a physical, gapped area with wooden planks. Then they tested their control scheme using the mini cheetah.
It was great fun working with a robot designed at MIT by some of our collaborators. Margolis states that the mini cheetah platform is great because it can be modified and consists of parts you can order online. If we needed a new camera or battery, it was easy to collect it from a regular supplier and install it with some help from Sangbae.” Margolis adds.
Sometimes, it wasn’t easy to estimate the robot’s current state. Real-world sensors are subject to noise, which can affect the outcome. This is not possible in simulation. Researchers used a motion capture system for specific experiments that required precise foot placement.
Their system performed better than those that used only one controller. The mini cheetah was able to cross 90% of the terrain.
“Our system adjusts the robot’s gait. This is a novelty. A human might try to jump across a wide gap by running fast. Then they might use their speed to gain speed. Next, they might connect both feet to leap across the gap. Margolis also says that robots can alter the timing and duration of foot contact to navigate terrain better.
Leaping out from the lab
Margolis states that while the researchers demonstrated that the control scheme works in a laboratory setting, there is still much to be done before the system can be deployed.
They plan to attach a giant computer to the robot in the future so that it can perform all of its computations onboard. They also plan to improve the robot’s state estimator to eliminate the need to use the motion capture system. They also want to enhance the low-level controller’s ability to exploit the robot’s full motion range and make the high-level controller more adaptable to different lighting conditions.
It is fantastic to see the flexibility of machine-learning techniques that can bypass carefully designed intermediate processes (e.g., Kim says that machine-learning approaches can forget centuries-old models and use state estimation and trajectory planning. “I’m excited about the future mobile robots equipped with robust vision processing specifically designed for locomotion.”
