Model Predictive Control of an Underdamped, Pneumatically Actuated, Soft Robot with Flexible Links for Unmodeled Environments

We start by listing our original three objectives for this work. For each objective, we then include a brief paragraph that summarizes what we have accomplished relative to that objective over the course of the program. 
1. Develop optimal control methods (Model Predictive Control) for underdamped, underactuated, soft robots that allow fast and precise motion despite the difficult natural dynamics of the system. 
2. Apply these control methods to achieve unprecedented performance in real and unmodeled environments for underdamped robot manipulation and locomotion. 
3. Test the controllers we develop in realistic and useful scenarios (such as equipment maintenance or installation, or exploration of rugged terrain), that will have wide applicability to future space missions as well as other crosscutting domains such as natural disaster relief or search and rescue missions. 
Objective 1: 
The field of soft robotics is a nascent field that is growing rapidly, but has mostly been focused in the areas of materials and construction of soft robot hardware. Over the past three years, we have instead focused on lumped parameter modeling of soft robots based on first principles and based on machine learning to enable model-based control. With the models that we have developed, we have shown that we can control large-scale soft (fabric-based) pneumatically actuated robots to commanded positions within a few millimeters. Additionally, because the joints for the specific hardware we have used are antagonistic, we have been able to show that we can control joint angle to within 1-2 degrees while simultaneously controlling the joint stiffness. In all cases, we have used different forms of Model Predictive Control (solving with both off-the-shelf optimization routines as well new implementations that take advantage of parallel computation with GPUs). However, we have also recently implemented other stochastic-based control methods such as Model Predictive Path Integral control in order to compare its performance to our own work for soft robot control. 
Objective 2: 
In order to enable soft robots to effectively complete real-world tasks one major hurdle that we had to address, and that was not originally in our proposed work, was state estimation. We addressed this problem by incorporating both optical-based sensors and algorithms as well as tendon-like integrated hardware with a learned calibration for real-world motion. In addition, because our soft robots are lightweight and should be easily transportable, we developed novel mobile manipulation kinematic planners that simultaneously solved for joint angles that would track a Cartesian trajectory as well as the position where the arm should be placed. We also developed new methods for designing soft robot manipulators based on specific soft robot metrics and showed that this method worked for designing manipulators, and more recently for designing the legs of a quadruped continuum soft robot. In fact, we showed that robots that were designed using effective metrics also made it easier to generate effective plans or trajectories for a given task. Finally, because we expect the dynamic model of our soft robot arms to change over time (both due to wear and tear and as they simply pick up a load), we developed a model reference adaptive controller that meshes with our existing model predictive controllers in order to adapt our dynamic models online as the soft manipulator performs pick and place tasks. 
Objective 3: 
Our planning code was effective and useful enough that our industrial partner who provided our soft robot hardware incorporated it into their code base and have been using it extensively. Our design optimization work from objective 2 was used to generate a design for a specific soft robot arm that was over six feet in length. This arm was built through an SBIR Phase II program with NASA and our collaborator Pneubotics. The arm was then attached to the planetary rover K-REX at NASA Ames. This culminated in testing the actual tasks that the robot was designed for (picking up rock samples from the ground, and making high velocity contact with the ground or rubble on the ground). Our eventual goal for the soft robot manipulators, in addition to servicing satellites or space habitats, would be for the robots to effectively and safely interact with humans (as astronauts or in other professions). As a first step towards this, we showed that using previously collected data of human-human dyads performing co-manipulation tasks with large objects, we could train a recursive neural net to allow a robot to predict human intent for the object when a robot and a human were instead working together. We believe that this is another area where soft robots can excel in the future as we develop inherently safer robots that can intuitively work with and assist human collaborators.