Non-linear control strategies for musculoskeletal robots

Recently, focus has shifted to more human-friendly robots, especially in the field of service or rehabilitation robotics, where research aims at bringing robots into increasingly unstructured environments. Here, humans still outperform robots in almost every aspect. One way towards this goal is to mimic more and more of the mechanisms of the human musculoskeletal system. These so called musculoskeletal robots incorporate compliant antagonistic actuation in a biologically inspired skeleton. Furthermore, complex joints like spherical joints and multi-articular muscles are realized. In the last decade impressive musculoskeletal robots have been developed, where the focus was mainly on constructing working prototypes. However, the field of controlling such robots has made very little progress, so far. Reasons for this are manifold, starting from unavailable sensor modalities, due to novel skeletal structures, to the difficulties in modeling the interaction of the muscles with the skeleton. In this work, (1) a generic model for the class of musculoskeletal robots was derived by extending standard models for robot dynamics, (2) several non-linear control strategies were developed and (3) a robotic arm was constructed to evaluate the novel control approaches. Non-linear controllers derived in this work include feedback linearization, which is also the main control technique used for tendon-driven robots and was therefore extended to be applied to the characteristics of musculoskeletal robots. Furthermore, a novel control law, based on Dynamic Surface Control (DSC), which is an extension to backstepping, was developed. In contrast to previously used techniques, the systematic approach of backstepping provides a method to obtain a cascade of controllers without neglecting the dynamics of the low-level control laws, hence providing provably stable closed-loop behavior. This scheme was further extended by adaptive neural network control to compensate for unmodeled dynamics such as friction. The dynamic model, incorporated by the different controllers utilizes a standard rigid body model of the skeleton which was extended to support spherical joints. In the presence of complex joint types and muscles that wrap around skeletal structures, analytic models of the mapping between the joint and the muscle space are extremely complex to obtain. To overcome this issue machine learning techniques were applied to attain the muscle kinematics. The different control laws were evaluated on a physical robotic platform which was specifically developed to cover the control challenges of the class of musculoskeletal robots. Hence it comprises a spherical shoulder joint and a revolute elbow joint. Actuation was realized by several uni-articular muscles and a bi-articular muscle spanning both joints. Results obtained on this platform show that it is possible to achieve stable closed-loop behavior, where the novel control laws developed in this work yield improved trajectory tracking performance, compared to existing approaches.