The skateboard rolled forward, picking up speed as the four-legged robot perched atop it glided along, dragging a small cart behind. Pilotless and joystick-less. The robot had learned the motion itself—balancing, pushing, even boarding—all thanks to a new form of artificial intelligence.
At the University of Michigan’s Computational Autonomy and Robotics Laboratory (CURLY Lab), computer scientists teamed up with engineers from the Southern University of Science and Technology in China to crack a problem that has long stumped robotics: how to get a machine to perform a task that involves continuous motion, abrupt changes, and direct contact with the world. Like skateboarding.
Their answer: a new framework called discrete-time hybrid automata learning, or DHAL.
Beyond Simple Walks
Robots have been walking, running, and even doing backflips for years. DHAL, however, changes how it deals with contact. Tasks like skateboarding involve smooth glides punctuated by sudden, discrete changes: stepping onto the board, pushing off, and shifting balance. This hybrid nature makes it difficult for robots to learn using conventional algorithms.
“Existing quadrupedal locomotion approaches do not consider contact-rich interaction with objectives, such as skateboarding,” Sangli Teng, the study’s lead author told Tech Xplore. “Our work was aimed at designing a pipeline for such contact-guided tasks that are worth studying, including skateboarding.”
Model-based approaches, which rely on equations and careful planning, often assume the robot will move in predictable ways—an assumption that crumbles on a moving skateboard. Model-free reinforcement learning, on the other hand, learns through trial and error but can’t easily recognize why or when a robot should change its behavior.
DHAL tries to get the best of both worlds. It learns distinct modes—like pushing and gliding—without needing a human to label when those modes begin or end. It also learns the smooth transitions between them, capturing the nuances of contact-heavy motion.
A Robot Learns to Ride
To test DHAL, the researchers chose a challenge that’s both visually striking and technically demanding: teaching a Unitree Go1 quadruped robot to skateboard. Inspired by real dogs that learn to ride boards, they designed a scenario that involved both propulsion and balance, on surfaces ranging from soft carpet to sloped pavement.
The robot wasn’t given step-by-step instructions. Instead, DHAL allowed it to discover how to move by learning the physics of each mode—when to push, when to stay still, and how to balance. This was achieved through a combination of a multi-critic learning system and a Beta distribution policy, which helped the robot make bounded, realistic movements without overshooting.
“Compared to the existing methods, DHAL does not require manual identification of the discrete transition or prior knowledge of the number of the transition states,” Teng said. “Everything in DHAL is heuristic and we showed that our method can autonomously identify the mode transition of dynamics.”
Unlike other learning methods that often try to brute-force their way to success by trying everything, DHAL provides structure. The result: a robot that could actually skateboard, not just wiggle in place.
Three Modes, One Flow
In experiments, the robot’s behavior naturally split into three modes: a pushing phase, a gliding phase, and an airborne transition. Using a built-in “automaton,” the robot learned to recognize which mode it was in at any given moment—without needing manual segmentation of the data.
“Even in the absence of external inputs, the robot could predict and adjust its trajectory with surprising accuracy,” the researchers wrote.
They visualized the robot’s internal decision-making using colored lights: green for mode one, blue for mode two, and red for mode three. In real-world tests, the robot succeeded in 100% of attempts on ceramic and carpeted floors—even when carrying extra weight or facing disturbances. It even handled slopes and small steps, albeit with lower success rates. In comparison, baseline methods without DHAL failed completely.
Why This Matters
Though the sight of a robot skateboarding may seem whimsical, the implications stretch far beyond. From rescue robots navigating rubble to warehouse bots pushing carts, many real-world systems must handle contact-rich environments and switch between different styles of movement.
The researchers note that current systems either rely on manually programmed rules or learn with little understanding of why a behavior works. DHAL adds a layer of intelligence that could enable more autonomous, adaptable, and safe robots in unpredictable settings.
Still, the system isn’t perfect. The researchers admit that it doesn’t yet generalize to highly dynamic skateboarding tricks—like the famed “ollie”—and more advanced perception systems will be needed to handle complex environments without relying on fixed connections between the robot and its skateboard.
But as a proof of concept, it’s a remarkable step.
