Researchers at ETH Zurich have developed a hybrid robot that combines the features of wheeled and legged robots, using reinforcement learning techniques.
Autonomous mobile robots could speed up goods delivery across different locations, addressing disruptions in supply chains. However, wheeled or legged robots may sometimes be enough to ensure efficient and independent deliveries.
Researchers at ETH Zurich’s Robotic Systems Lab have introduced a hybrid robot that merges the features of wheeled and legged robots. The robot employs reinforcement learning techniques to switch seamlessly between driving and walking, adapting to various terrains.
The robotic system developed by the team improves upon a previous design from team CERBERUS, which included members from the indoor drone company Flyability and won the DARPA Subterranean Challenge in 2021. Unlike the CERBERUS robot, their new model features a simplified design and a more sophisticated AI-powered navigation system.
The researchers equipped their robot with the ability to navigate autonomously by developing, training, and testing various hierarchical reinforcement learning techniques. They ultimately created a neural network-based controller capable of processing diverse inputs and quickly generating new navigation strategies for the robot within milliseconds.
The ETH Zurich-developed robot drives using its wheels on easy, smooth terrain, which helps conserve energy. When encountering complex terrains that are challenging or impossible for wheels alone, such as steps, it switches to walking mode.
The neural network-based controller, developed and trained by the team, processes sensory data to identify the most efficient travel method for different terrains. This integration allows the robot to utilize the advantages of both wheeled and legged robots effectively.
The controller designed by the team does not use traditional planning and model-based control techniques. These conventional methods often struggle in real-world environments filled with uncertainty and random disturbances.
Reference: Learning robust autonomous navigation and locomotion for wheeled-legged robots. Science Robotics(2024). DOI: 10.1126/scirobotics.adi9641.
