Robotics is where AI meets the real world of industrial applications. ANYbotics uses AI to improve its robot’s locomotion, navigation, and inspection intelligence. This video and article provide insight into how, through deep reinforcement learning, ANYmal learns to operate safely and effectively in complex terrain, such as climbing open-riser grated stairs.
Robot Locomotion: Optimization and Learning Approaches for Walking
For complex systems, such as legged robots, there are two basic approaches to motion planning and control:
- Optimization-based approach to identify the desired locomotion behavior online
- Deep reinforcement learning-based approach to train the model in offline simulation
ANYmal uses reinforcement learning to teach itself to walk as toddlers do by trial and error. However, unlike humans, ANYmal’s learning journey happens entirely in a simulation before deploying it in the physical world.
In their work ScienceRobotics, Hwangbo Jemin and Lee Joonho, and their colleagues at ETH Zurich convincingly demonstrated how reinforcement learning can be applied to legged robots. ANYbotics engineers successfully built on their work to upgrade ANYmal’s locomotion, a solution we call Trekker. The result is an ANYmal with next-level, industry-grade locomotion, capable of significantly outperforming optimization-based approaches in terms of robustness in challenging environments.
In the simulation, ANYbotics provides ANYmal with challenges like terrain, efficiency, posture, and hardware limitations. Within minutes and hours, the simulated ANYmal undergoes millions of variations, finally converging to a model of the best locomotion patterns. The best model is transferred as a new capability to the physical robot – where it undergoes rigorous testing before being rolled out to the robotic fleet as an updated version.
Robust Mobility in the Industrial World
The learning-based approach proved to result in more capable and robust locomotion patterns of ANYmal for its primary application in complex industrial facilities. Another advantage is that new capabilities can be deployed quickly across all operational robots after being optimized in the simulation. For example, the recently improved near-field perception made the robot safely negotiate steep, open grated stairs. While these stairs are commonly found on industrial sites, it is a cutting-edge attainment for legged robots.
Furthermore, in research scenarios, these robots are trained to perform extreme movements like parkour, showcasing their agility and adaptability in dynamic environments. Using reinforcement learning-based control, these robots overcome very complex obstacles. Thereby, the robot can not only walk, but it can literally use its main body and knees to overcome large obstacles like boxes.
Using AI to Drive Continued Innovation
Our job at ANYbotics is to apply what we learn to our autonomous workforce and solve real-world problems better than anyone else. With an AI-centric approach, our robots have an advantage in solving these problems and are better prepared to face unknown challenges. But that’s not all. We are committed to continuous improvement and innovation by infusing our research and development with AI techniques. This is how we ensure we are at the forefront of shaping the future of automation.