Title: "Evolving bio-inspired controllers with once per cycle feedback for robust dynamic walking"
Abstract:
In recent years there has been a growing interest in the field of bio-inspired robots and dynamic walking. This research field has a wide range of applications including search and rescue robots, human exoskeletons and gait rehabilitation. However, while walking and running on a flat surface have been studied extensively, walking dynamically over terrains with varying slope remains a challenge. In the first part of my research I designed a feed forward (FF) controller based on a Central Pattern Generator (CPG) that achieved stable dynamic gaits on a simulated compass biped (CB) model. Applying short torque pulses, instead of enforcing joint trajectories, enabled the model to move according to its natural dynamics. The controller coordinated the activation of torque pulses and achieved stable gaits over a limited range of slopes.
In the second part of my research I focused on improving the robustness to slope variations by applying a once per cycle feedback to the CPG controller. The terrain's slope is measured and used to adapt the FF controller to the terrain, by modifying both the CPG frequency and the torque amplitude once per step. This approach maintains the advantage of the FF controller in exploiting the system’s natural dynamics without overtaxing the requirements on update rate and measurement fidelity. In order to optimize the controller’s parameters and generate stable gaits with the largest robustness to slope variations, multi-objective optimization (MOO) was performed using a genetic algorithm (GA). The algorithm evolved populations of controllers by evaluating their performance in walking speed, energy efficiency, rate of convergence to a limit cycle and range of stable slopes. The algorithm generated controllers with distinct abilities which can be roughly divided into four groups: (I) fast walkers, (II) efficient walkers, (III) quickly converging walkers and (IV) good climbers. The best climber successfully traversed terrains with slopes ranging from +7 to -8 degrees, comparable to most slopes found in human constructed environments, relying on the minimal feedback strategy. Gait stability was verified by computing the linearized Poincare Map both numerically and theoretically.
In parallel, I also was part of ROBIL – the Israeli group that participated in the DARPA Virtual Robotics Challenge. As part of the team I developed a bio-inspired quadruped gait for the humanoid robot ATLAS. This quadruped gait enabled ATLAS to complete the dismounted mobility part of the challenge which included crossing a mud pit, a hilled area and an area with debris. I will briefly describe the results of this approach.
Социальные закладки