Researchers create robot with wheels and legs that navigates autonomously

Autonomous and fast mobile robots could help deliver goods to various locations, helping to combat disruptions in product supply chains. However, wheeled or legged robots alone may not be enough to deliver efficiently and independently.

Researchers from the Robotic Systems Lab at ETH Zurich recently presented a new robot design combining the capabilities of wheeled and legged robots. This robot, presented in a Scientific robotics paper, navigates environments using various reinforcement learning techniques, which allow it to smoothly switch between driving and walking modes, adapting to different terrains.

“The main goal of the project was to build a large-scale autonomous driving system for such a ground robot, with the fastest speed ever seen,” Joonho Lee, co-author of the paper, told Tech Xplore . “It is the result of more than five years of research into legged robotics, autonomous navigation and robot perception.”

The robotic system developed by Lee and his colleagues builds on a previous robot created by the CERBERUS team, a team including researchers from indoor drone company Flyability, which won the DARPA Subterranean Challenge in 2021. Unlike the robot developed by the CERBERUS team, however, their system has a simplified design and a more advanced navigation system powered by AI.

“Traditionally, navigation planning for ground robots was done using online optimization methods,” Lee explained. “Such approaches work well for single-wheeled robots or slow-walking robots, but in the case of fast robots like ours (which can travel up to 20 km/h), they cannot provide maps of sufficiently fast navigation. m/s, a delay of 0.5 seconds can cause an error of 1 m, which can lead to a catastrophic collision.

To enable their robot to autonomously navigate environments, the researchers developed, trained and tested various hierarchical reinforcement learning techniques. Ultimately, they trained a neural network-based controller that could process different types of inputs, creating new navigation plans for the robot in milliseconds.

“Another great advantage of our approach is that our neural network controller fully understands the nonlinear and complex dynamics of legged robots,” Lee said. “As it understands the behavior of the robot on different terrains and at different speeds, it can steer it very effectively.”

On smooth, easy-to-navigate terrain, the robot developed at ETH Zurich moves forward, thus using its wheels and minimizing energy consumption. In more complex terrain where it would be difficult or impossible to navigate using wheels, such as where steps are present, the robot can switch to walking mode.

The neural network-based controller developed and trained by Lee and his colleagues can process sensory data to determine the most efficient way for the robot to move over specific terrains. This allows the robot to effectively combine the strengths of conventional wheeled robots with those of legged robots.

“Wheeled robots are efficient but cannot overcome high obstacles,” Lee said. “On the other hand, legged robots are very effective in overcoming obstacles and steep slopes, but their efficiency is very low because they have to walk more than 10 joints irregularly. Usually, walking robots can only work for 1 hour maximum Thanks to its roller legs, our robot can overcome the same obstacles as normally walking robots with operation at least 3 times longer.”

The controller developed by Lee and colleagues does not use traditional model-based planning and control techniques. Notably, these traditional methods have often proven to be ineffective in real-world contexts characterized by uncertainty and random disturbances.

Instead, the team’s controller is driven by two artificial neural networks. These networks process the data collected by the sensors built into the robot, produce appropriate walking movements and decide in which direction the robot should move.

“To train a navigation agent, we created a special simulation environment, which resembles a computer game,” Lee explained. “Our software automatically generates new “steps” for the navigation controller with different terrain and complex disturbances. After several hours of training, we obtained very robust and versatile neural network controllers, capable of handling all kinds of rough terrain and maze-like environments.”

Another advantage of the navigation system controlling the robot’s movements is that it is simpler than many existing controllers. One of the two neural networks it relies on focuses on planning walking movements, while the other focuses on overall robot navigation. The controller also includes basic terrain mapping and SLAM (simultaneous localization and mapping) modules.

“This is the simplest navigation system design I have seen, while very powerful neural network controllers remove a lot of engineering effort in system integration,” Lee said. “The actual time we spent building the navigation system itself was less than a year.”

Lee and his colleagues tested their navigation system in a series of experiments carried out in real-world environments. They found it to be very responsive and high performing, as it allowed their robot to successfully travel over 10 km across two different European cities, namely Zurich and Seville.

In the future, the wheeled robot and navigation system presented in this recent article could be further improved and deployed in various contexts. One of their most promising applications will be the fast, reliable and autonomous delivery of goods across different terrains.

“I now want to expand this system with multimodal inputs,” Lee added. “Currently it only relies on geometric information for navigation and walking, but in the real world we need to consider more things when we walk around. For example, the robot should care about more information semantics, like checking if the ground is wet, if it should stay on the sidewalk or in the grass, if a red light is on, etc.

© 2024 Science X Network