An Integrated Concept for Footstep Planning and Navigation for Different Types of Multi-Legged Robots in Challenging Environments

Among a large variety of available mobile ground robotic designs, legged robots are potentially the most versatile systems with respect to the traversal of uneven ground. However, they also have to deal with the most complex trajectory planning and control problems, when compared, e.g., to ground robots locomoting on wheels or tracks. The ability of legged robots to generate continuous motion over larger distances of uneven terrain by using discrete foot displacements leads to especially capable locomotion abilities in challenging environments. This potentially allows them to cross a large variety of non-contiguous surfaces (e.g. stepping stones) and to climb over obstacles (e.g. debris). In recent years, interest in humanoid and four-legged robots has significantly increased due to the attention and efforts initiated through the DARPA Robotics Challenge (2013-2015) leading to significant progress in legged robot hardware and motion controller performance. However, to fully benefit from the potential of these unique locomotion capabilities, advances in autonomous functionality is required as the high mechatronic complexity of such robots makes teleoperation at the joint-level by humans highly infeasible. The ability of supervision by a remote human operator on a high level of abstraction is needed to traverse previously unknown terrains with a legged robot for the first time. This thesis addresses several such advances: Efficient perception especially of uneven, rough and irregular terrain, and suitable digital representations of them are developed as mandatory prerequisites for developing new methods for finding feasible (and posturally stable) foothold positions and their alignment in such challenging terrain. On top of these, planning and navigation of sequences of suitable footsteps is developed to traverse a terrain section. Execution of footstep sequences over long distances requires methods using the available robot specific legged locomotion controllers for repeated online updating and stitching of footstep plans to achieve continuous autonomous walking over long distances in uneven terrain. Moreover, the reusability of methods for perception, planning, and step execution for different types of legged robots is highly desirable to efficiently adapt and deploy the resulting framework to different types of robots and their specific kinematic and dynamic motion properties as well as their motion controllers. The innovations, methods and their implementation presented in this work are successfully evaluated in challenging irregular terrain with seven different types of state of the art robots: two versions of Boston Dynamics' humanoid Atlas robot, two versions of the THORMANG robot by Robotis equipped with custom sensing and computing, the humanoid WALK-MAN robot, the quadruped ANYmal robot of ANYbotics and a tracked robot with four actuated flipper wheels (Telemax Hybrid by Telerob). It is successfully demonstrated how the novel footstep planning framework extends the current state of research in the field of 3D planning by applying the approaches developed in this thesis to improve the overall performance of the planner and efficiently generate suitable step plans even in previously not reasonable traversable irregular terrain. The generalizability and modularity of the proposed framework allows the flexible application with different types of legged robots but also, as demonstrated in the evaluation, generalizes beyond the intended primary application purpose, e.g. as automatic flipper control for a tracked robot crossing rough terrain