American engineers taught the two-legged Cassie robot to detect low obstacles and bend over to pass under them.
Walking robots are gradually emerging from laboratories to the market, but while their hardware is ready to walk in a complex environment, it is much better than software. As a rule, they are able to overcome small obstacles at the level of their feet, but are unable to take into account obstacles that limit movement not from below, but from above. In the summer, a group of engineers from the University of California, Berkeley, led by Koushil Sreenath, introduced a motion planner for four-legged robots, which allows them to independently assess objects in front of them and jump if their height is not too high, and subject to restrictions from above.
In the new work, they solved a slightly different problem and on a fundamentally different platform: passage under obstacles using the bipedal robot Cassie. Since it uses a dynamic gait, this task is more difficult than on the initially stable four-legged robots. The engineers described the robot as a spring-loaded reverse pendulum. This type of model has already been used in other robots, for example, Salto, but the authors of the new work added a variable height of the pendulum to the model.
A depth camera is installed in the upper part of the bipedal robot Cassie, with the help of which it marks the space in front of it on voxels – a three-dimensional analogue of pixels. On the basis of a three-dimensional map of obstacles, the robot also builds a two-dimensional one, composed of pixels with a size of 0.5 meters. Depending on the filling of voxels, the map is marked into free and occupied areas, as well as those in which there is a surmountable obstacle. Such an obstacle should be located at a height of 0.7 to 1 meter – this is the minimum and maximum height of the robot.
During the movement, the robot is controlled by three planners. The global receives from a person the end point to which it is necessary to arrive, and builds a simplified route using a 2D map. The local receives intermediate points from the global and builds a trajectory. And the reactive planner is responsible for the walk itself, considering only the next 30 centimeters of the path and giving parameters such as speed and altitude to the walk controller.
The authors showed a video with several tests. While the robot walks many times slower than it is technically capable, however, it has really learned to assess obstacles on the way and adapt its height to them.
Cassie’s creators from Agility Robotics recently showed the maximum potential of bipedal robot Cassie. During the tests, he was able to go 5 kilometers in 44 minutes, never falling and reaching a speed of 2.15 meters per second on one of the circles.