Legged robots skateboard successfully with reinforcement learning framework

Legged robots, which are often inspired by animals and insects, could help humans to complete various real-world tasks, for instance delivering parcels or monitoring specific environments. In recent years, computer scientists have created algorithms that allow these robots to walk at different speeds, jump, emulate some of the movements of animals and move with great agility.

Researchers at the University of Michigan’s Computational Autonomy and Robotics Laboratory (CURLY Lab) and Southern University of Science and Technology have now developed a reinforcement learning-based framework that allows legged robots to use a skateboard successfully. This framework, outlined in a paper on the arXiv preprint server, could also be used to emulate other real-world complex movements that entail physical contact with nearby objects.

“Existing quadrupedal locomotion approaches do not consider contact-rich interaction with objectives, such as skateboarding,” Sangli Teng, corresponding author of the paper, told Tech Xplore. “Our work was aimed at designing a pipeline for such contact-guided tasks that are worth studying, including skateboarding. The University of Michigan has a long history of developing hybrid dynamical systems, which inspired us to identify such hybrid effects via data-driven approaches in AI.”

The main goal of the recent work by Teng and his colleagues was to allow legged robots to perform contact-guided motions, including skateboarding. To achieve this, they developed a new framework called discrete-time hybrid automata learning (DHAL).

“Hybrid dynamics” means a system can perform both continuous and discrete state transitions. This essentially means it can move smoothly and suddenly change its state over time.

“For example, when a bouncing ball interacts with the ground, the ball has continuous dynamics in the air and discrete state transitions when colliding with the ground,” explained Teng.

“For systems with multiple continuous dynamics and transition functions, it is extremely difficult to identify the discrete mode and continuous dynamics at the same time. This is because a possible transition grows exponentially fast with regards to the number of possible discrete transitions.”

The abrupt transitions described by Teng make it difficult for conventional regression-based computational methods to learn the dynamics of a system. DHAL, the framework developed by the researchers, can identify these sudden transitions, subsequently learning each continuous segment of a system’s dynamics using regression-based techniques, reducing the discontinuous effect that was found to impair the performance of robots on tasks such as skateboarding.

“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,” said Teng. “Everything in DHAL is heuristic and we showed that our method can autonomously identify the mode transition of dynamics.”

A further advantage of the DHAL framework is that it is highly intuitive, thus ensuring that the mode transitions it identifies are aligned with those typically associated with skateboarding. In initial tests, the researchers found that it allowed four-legged (i.e., quadruped) robots to smoothly step onto a skateboard and use it to rapidly move forward while also pulling a small cart behind them.

“In the pushing, gliding and upboarding phase, DHAL will automatically output different labels,” said Teng. “Our method can be applied to state estimation of hybrid dynamical systems to find out if such transition occurs. With this transition information, the system can better estimate the states to assist the decision making.”

The new reinforcement learning framework Teng and his colleagues developed could soon open new possibilities for the real-world deployment of legged robots. For instance, it could allow them to move faster using a skateboard, delivering packages across urban environments, inside offices or manufacturing facilities.

“We now plan to apply this framework to other scenarios, such as dexterous manipulation (i.e., the manipulation of objects with multiple fingers or arms),” added Teng. “DHAL is expected to predict the contact more accurately, thus allowing planning and control algorithms to make better decisions.”

© 2025 Science X Network