Real-Time Optimization for Control of a Multi-Modal Legged Robotic System
For decades humans have been trying to create machines that can mimic the capabilities of legged animals. Already humans have developed cars capable of traveling long distances at high speeds, and automation robots able to perform incredibly accurate tasks at high rates. However, the machines and robots that we have created so far only operate in specific controlled environments. What is impressive about animals is their capability of navigating a multitude of environments. Further still, humans and select animals are not only able to navigate the world but they are also capable of manipulating the world around them. It is these abilities that we wish to replicate in robots in order to create machines that can aid humans in their everyday lives.
In recent years, there has been a large surge in the number of quadrupedal robots available for commercial use. They have shown that they can navigate the world at speeds comparable to human walking or jogging. However, as of now no quadrupedal robot has the capability to help with tasks other than search or surveillance type of applications. Humanoid robots have also made large advancements in the past decade. Multiple full sized humanoids have demonstrated stable walking and locomotion. However, the most important development for humanoid robots is their ability to manipulate their environment. The DARPA Robotics Challenge show cased humanoid robots driving cars, opening doors, and using power tools to complete tasks. Unfortunately, right now there are no commercially available humanoid robots due in most part to their lack of safety and reliability. Unlike quadrupedal robots, humanoids are not statically stable and thus require active balancing at all times. This added complexity makes humanoid robots dangerous to operate in human environments.
This dissertation presents a novel quadrupedal robot and control strategy that brings legged robotics closer to one day aiding humans in their everyday life. The novel quadrupedal robot, named ALPHRED, uses a non-traditional kinematic structure to provide multiple modes of operation, giving the robot capabilities beyond that of traditional quadrupeds. The online path planner was developed using a non-linear program to solve for viable center of mass trajectories and footstep locations based off of the dynamics of the robot and height map information from vision data. In addition, a model predictive controller was developed to track the desired motion of the robot. This model predictive controller uses massless leg dynamics to predict the dynamics of the body but also incorporates known swing leg dynamics in a novel fashion that results in a 250 Hz motion tracking controller. Using the developed control schemes the robot is capable of navigating uneven terrain and can achieve a top speed of 1.5 m/s on flat terrain. ALPHRED is equipped with passive wheels on it's belly that allow it to push itself around on flat ground for an efficient and fast form of locomotion. Finally, ALPHRED is capable of balancing on two limbs freeing up the other two limbs for manipulation tasks. Using all of the mode's of operation ALPHRED is the first quadrupedal robot that is capable of picking up varying size packages and completing an end-to-end package delivery.