Kit Based Motion Generator for a Soft Walking Robot

To control the motion of legged robots, a reference must be generated for each individual actuator. This can be done either off-line, by trajectory optimization, or on-line by means of model predictive control. If the motion patterns are to be generated on-line, it must be ensured that the algorithms used are sufficiently fast, since the robot has to continuously react to the current state in order to achieve robust behaviour [1].

However, this only applies to dynamic motion. Provided that a robot always maintains quasi-static equilibrium, the runtime of the motion generating algorithms is irrelevant for the robustness of the motion, but essential for a smooth motion sequence.

In [2], we presented a gecko-inspired soft robot that is able to move in any direction. Until now, the robot was operated in joint space (bending angles of the individual limbs) with predefined gait patterns. A closed loop control to be able to target specific references in the task space (Cartesian coordinates of the robot) was previously not possible due to a missing mathematical transformation from task to joint space.

The approach presented here selects the optimal gait pattern from a discrete, predefined set of possibilities to get closer to a given target position; similar to [3]. The method is based on an off-line component: elementary gait patterns are generated by trajectory optimization using the simulation model from [4]. And an on-line component: for given robot and target positions the optimal next elementary gait pattern is chosen based on a minimization problem, and the joint space references are derived from it.

Since the robot holds on to the platform with suction cups, i.e. its feet cannot be moved in the fixed state, it cannot transition to any subsequent pose from its current one. To ensure feasible subsequent poses, the elementary patterns always begin and end with one and the same pose, so that they can be placed on top of each other like Lego bricks. Whenever more than one choice is available as a subsequent pose, the robot position and the target position are measured and the optimal subsequent pose is selected. In this way, modelling errors can be compensated.

A great advantage of this method is a straightforward transition between different motion modes, such as from trotting to crawling. It is discussed how many different elementary patterns there must be to ensure stable position control. Finally, we show in simulation and experiment that the robot can master arbitrary obstacle courses by making use of the proposed motion generator.

[1] Wieber, Pierre-Brice, Russ Tedrake, and Scott Kuindersma. "Modeling and control of legged robots." Springer handbook of robotics. Springer, Cham, 2016. 1203-1234.

[2] Schiller, Lars, Arthur Seibel, and Josef Schlattmann. "Toward a Gecko-Inspired, Climbing Soft Robot." Frontiers in Neurorobotics 13 (2019): 106.

[3] Kuffner, James, et al. "Motion planning for humanoid robots." Robotics Research. The Eleventh International Symposium. Springer, Berlin, Heidelberg, 2005.

[4] Schiller, Lars, Arthur Seibel, and Josef Schlattmann. "A Lightweight Simulation Model for Soft Robot's Locomotion and its Application to Trajectory Optimization." IEEE Robotics and Automation Letters 5.2 (2020): 1199-1206.