ANYmal Robot Shows Superhuman Off-Road Mobility
Robots, like humans, usually rely on two different sensory modalities when interacting with the world. There is exteroceptive perception (or exteroception), which comes from external sensory systems such as lidar, cameras and eyes.
There is also proprioceptive perception (or proprioception), which is internal sensing that includes things like touch and sensing force. As a rule, we humans use both of these sensory modalities at the same time to move around, with exteroception helping us plan ahead and proprioception kicking in when difficulties arise.
You use proprioception, for example, in the dark, when movement is still quite possible — you simply do it slowly and carefully, relying on balance and touch.
For robots, exteroception is what lets them do all the cool things: with really good external perception and time (and computing power) to plan amazing movements, robots can move dynamically and quickly.
However, legged robots are much less capable in the dark, or under any circumstances where the exteroception they need either fails (because a sensor is not working for some reason) or simply performs poorly due to robot-unfriendly things such as reflective surfaces, dense undergrowth or other obstacles. This is a problem, since the real world is full of robot-unfriendly things.
A study published by the Robotic Systems Lab at ETH Zurich in Science Robotics describes a control system that allows a robot to assess how reliable the exteroceptive information it receives is. When the data is good, the robot plans ahead and moves quickly.
But when the dataset seems incomplete, noisy or misleading, the controller gracefully degrades to proprioceptive locomotion. This means the robot keeps moving — perhaps more slowly and cautiously, but it keeps moving — and eventually it reaches a point where it can again rely on exteroceptive perception.
This is a technique that people and animals use, and now robots can use it too, combining speed and efficiency with safety and reliability to handle virtually any difficult terrain.
Everyone got an exciting preview of this technique during the final round of the DARPA SubT competition last fall, when the legged ANYmal robots of Team Cerberus managed to win. The performance in the video below is impressive, especially considering that ANYmal completed the uphill climb four minutes faster than the recommended time for an average human.
The footage of the ANYmal robot walking through dense vegetation and deep snow clearly shows how well the system works. While the exteroceptive data shows obstacles everywhere and inaccurate ground height, the robot knows where its feet are and relies on this proprioceptive data to move forward safely without falling.
Other legged robots use proprioception for reliable movement, but what is unique here is the smooth combination of speed and reliability, as the controller switches between exteroception and proprioception depending on how confident it is in what it sees. The results of ANYmal on this hike, as well as during the SubT final, are convincing proof of how well this approach works.
For more details, IEEE Spectrum spoke with the paper's author, Takahiro Miki, a PhD student at the Robotic Systems Lab at ETH Zurich.
“Achieving animal-like mobility is probably a goal for many of us researchers in this field. However, robots still lag far behind nature, and this paper is only a tiny step in that direction. <…> We aim to send robots to places that are too dangerous or difficult for humans. In this work, by "harsh" we mean places that are tough for us, not just for robots. For example, steep hiking trails or snowy trails that are hard to walk on. With our approach, the robot traversed steep and wet rocky surfaces, dense vegetation, or rough terrain in underground tunnels and natural caves with loose gravel at human speed”
According to him, the intuition of the human operator choosing a possible path for the robot certainly helped the robot succeed. Although the robot is sturdy, it cannot overcome obstacles that are physically impossible, such as obstacles larger than the robot or cliffs.
However, in other scenarios, such as the DARPA SubT Challenge, the robot is controlled by a high-level exploration and path planning algorithm. This planner knows the capabilities of the motion controller and uses geometric cues to control the robot safely. Achieving this for an autonomous hike in mountainous terrain, where a more semantic understanding of the environment is needed, is future work for the scientists.
The snowy staircase was the very first outdoor experiment the scientists conducted with the current controller, and Miki was surprised that the robot could handle the slippery snowy stairs. The terrain during the hike was also quite steep and challenging. When the PhD student first checked the terrain, he thought it might be too difficult for the robot, but it overcame all the obstacles.
“We think the robot determines whether exteroception matches proprioception through foot contact or foot position. If the map is correct, the feet touch where the map suggests. The controller then recognizes that exteroception is correct and uses it. As soon as it senses that foot contact does not match the ground on the map or the feet go below the map, it recognizes that exteroception is unreliable and relies more on proprioception. We demonstrated this in an additional video experiment”
To date, the scientists have controlled the robot manually in most experiments (except the DARPA SubT Challenge). The next goal is to add more levels of autonomy. Miki wants the robot to complete a complex hike without human involvement.
In addition, there is great potential to improve the robot's mobility. For "harsher" landscapes, the researchers want the robot to perceive the world in 3D and exhibit more varied behavior, such as jumping over steps or crawling under overhanging obstacles, which is impossible with the current 2.5D height map.
