Watching a cheetah sprint across uneven terrain with graceful leaps appears effortless, but replicating this agility in robotic systems presents significant engineering challenges.
While quadrupedal robots inspired by animal locomotion have advanced tremendously in recent years, they still struggle when navigating landscapes with sudden elevation changes and gaps that their biological counterparts traverse with ease.
"In challenging environments, vision becomes essential for preventing failures. Without visual perception, avoiding obstacles like gaps becomes nearly impossible. Existing approaches for integrating vision with legged locomotion often lack compatibility with next-generation agile robotic platforms," explains Gabriel Margolis, a PhD researcher at MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) under Professor Pulkit Agrawal.
Margolis and his research team have pioneered a revolutionary system that dramatically enhances the speed and agility of legged robots when jumping across terrain gaps. This innovative control architecture features two complementary components—one that processes real-time visual input from a front-mounted camera and another that translates this data into precise movement instructions. The team validated their approach using MIT's mini cheetah, a highly capable and agile robot developed in Professor Sangbae Kim's mechanical engineering laboratory.
Unlike conventional quadruped robot control methods, this dual-component system eliminates the need for pre-mapped terrain, enabling unprecedented autonomy. This breakthrough capability could eventually allow robots to navigate forested areas during emergency response missions or ascend staircases to deliver medication to homebound elderly individuals.
Margolis authored the research paper alongside lead investigator Pulkit Agrawal, who directs MIT's Improbable AI Lab and serves as the Steven G. and Renee Finn Career Development Assistant Professor in the Department of Electrical Engineering and Computer Science. Additional contributors include MIT Professor Sangbae Kim from mechanical engineering, graduate students Tao Chen and Xiang Fu, Arizona State University graduate student Kartik Paigwar, and University of Massachusetts at Amherst assistant professor Donghyun Kim. The findings will be presented at the upcoming Conference on Robot Learning.
Advanced Control Architecture
The innovation of this system lies in its dual-controller approach working in harmony.
A controller functions as an algorithm that converts the robot's current state into actionable movement commands. Many existing blind controllers—those without visual input—offer robustness but only enable navigation across continuous surfaces.
Vision presents such complex sensory data that traditional algorithms struggle to process it efficiently. Most vision-based systems rely on terrain "heightmaps" that must be either pre-constructed or generated in real-time, a process that typically proves slow and failure-prone when the heightmap contains inaccuracies.
The researchers developed their system by integrating the most effective elements from robust blind controllers with a separate module that handles vision processing in real-time.
The robot's camera captures depth images of approaching terrain, feeding this information along with the robot's physical state data (joint angles, body orientation, etc.) to a high-level controller. This controller utilizes a sophisticated neural network that "learns" through experience.
This neural network generates a target trajectory, which the second controller converts into precise torque calculations for each of the robot's 12 joints. Unlike the high-level controller, this low-level controller doesn't rely on neural networks but instead uses concise physical equations that accurately model the robot's motion dynamics.
"This hierarchical approach, particularly the implementation of the low-level controller, allows us to constrain the robot's behavior for more predictable operation. With this low-level controller, we can apply well-defined models with specific constraints—something typically challenging in learning-based networks," Margolis explains.
Training the Neural Network
The researchers employed reinforcement learning—a trial-and-error training method—to develop the high-level controller. They conducted hundreds of simulations featuring the robot navigating diverse discontinuous terrains, rewarding successful crossings.
Over time, the algorithm learned to identify actions that maximized these rewards.
The team then constructed a physical test environment with gaps using wooden planks and evaluated their control system using the mini cheetah robot.
"Working with a robot developed in-house at MIT by our collaborators was incredibly rewarding. The mini cheetah provides an excellent experimental platform due to its modular design and construction from mostly off-the-shelf components. When we needed a new battery or camera, we could simply order it from standard suppliers and, with assistance from Sangbae's lab, integrate it into the system," Margolis notes.
Accurately estimating the robot's state presented challenges in real-world testing. Unlike simulations, physical sensors encounter noise that accumulates and affects performance. For experiments requiring precise foot placement, the researchers implemented a motion capture system to determine the robot's exact position.
Their system demonstrated superior performance compared to single-controller alternatives, with the mini cheetah successfully navigating 90% of the test terrains.
"Our system uniquely adjusts the robot's gait patterns. When humans need to leap across a wide gap, they typically build momentum by running quickly before executing a powerful two-footed jump. Similarly, our robot modifies the timing and duration of its foot contacts to optimize terrain traversal," Margolis describes.
Future Applications
While the researchers successfully demonstrated their control system in laboratory settings, significant development remains before real-world deployment becomes feasible, Margolis acknowledges.
The team plans to integrate a more powerful onboard computer to enable all processing to occur on the robot itself. They also aim to enhance the robot's state estimation capabilities to eliminate dependence on external motion capture systems. Additionally, they intend to improve the low-level controller to fully exploit the robot's range of motion and refine the high-level controller for reliable operation under varying lighting conditions.
"It's remarkable to observe how modern machine learning techniques can bypass meticulously designed intermediate processes—such as state estimation and trajectory planning—that traditional model-based approaches have depended on for centuries," Kim comments. "I'm enthusiastic about the future of mobile robots equipped with more robust vision processing systems specifically trained for locomotion tasks."
The research received support from MIT's Improbable AI Lab, Biomimetic Robotics Laboratory, NAVER LABS, and the DARPA Machine Common Sense Program.
