Design Optimization for Rough Terrain Traversal Using a Compliant, Continuum-Joint, Quadruped Robot.

Legged robots have the potential to cover terrain not accessible to wheel-based robots and vehicles. This makes them better suited to perform tasks such as search and rescue in real-world unstructured environments. In addition, pneumatically-actuated, compliant robots may be more suited than their rigid counterparts to real-world unstructured environments with humans where unintentional contact or impact may occur. In this work, we define design metrics for legged robots that evaluate their ability to traverse unstructured terrain, carry payloads, find stable footholds, and move in desired directions. These metrics are demonstrated and validated in a multi-objective design optimization of 10 variables for a 16 degree of freedom, pneumatically actuated, continuum joint quadruped. We also present and validate approximations to preserve numerical tractability for any similar high degree of freedom optimization problem. Finally, we show that the design trends uncovered by our optimization hold in two hardware experiments using robot legs with continuum joints that are built based on the optimization results.

Using First Principles for Deep Learning and Model-Based Control of Soft Robots.

Model-based optimal control of soft robots may enable compliant, underdamped platforms to operate in a repeatable fashion and effectively accomplish tasks that are otherwise impossible for soft robots. Unfortunately, developing accurate analytical dynamic models for soft robots is time-consuming, difficult, and error-prone. Deep learning presents an alternative modeling approach that only requires a time history of system inputs and system states, which can be easily measured or estimated. However, fully relying on empirical or learned models involves collecting large amounts of representative data from a soft robot in order to model the complex state space-a task which may not be feasible in many situations. Furthermore, the exclusive use of empirical models for model-based control can be dangerous if the model does not generalize well. To address these challenges, we propose a hybrid modeling approach that combines machine learning methods with an existing first-principles model in order to improve overall performance for a sampling-based non-linear model predictive controller. We validate this approach on a soft robot platform and demonstrate that performance improves by 52% on average when employing the combined model.

Model Reference Predictive Adaptive Control for Large-Scale Soft Robots.

Past work has shown model predictive control (MPC) to be an effective strategy for controlling continuum joint soft robots using basic lumped-parameter models. However, the inaccuracies of these models often mean that an integral control scheme must be combined with MPC. In this paper we present a novel dynamic model formulation for continuum joint soft robots that is more accurate than previous models yet remains tractable for fast MPC. This model is based on a piecewise constant curvature (PCC) assumption and a relatively new kinematic representation that allows for computationally efficient state prediction. However, due to the difficulty in determining model parameters (e.g., inertias, damping, and spring effects) as well as effects common in continuum joint soft robots (hysteresis, complex pressure dynamics, etc.), we submit that regardless of the model selected, most model-based controllers of continuum joint soft robots would benefit from online model adaptation. Therefore, in this paper we also present a form of adaptive model predictive control based on model reference adaptive control (MRAC). We show that like MRAC, model reference predictive adaptive control (MRPAC) is able to compensate for "parameter mismatch" such as unknown inertia values. Our experiments also show that like MPC, MRPAC is robust to "structure mismatch" such as unmodeled disturbance forces not represented in the form of the adaptive regressor model. Experiments in simulation and hardware show that MRPAC outperforms individual MPC and MRAC.