Adaptive neural network Dynamic Surface Control: An evaluation on the musculoskeletal robot Anthrob

Abstract

Abstract-The soft robotics approach is widely considered to enable robots in the near future to leave their cages and move freely in our modern homes and manufacturing sites. Musculoskeletal robots are such soft robots which feature passively compliant actuation, while leveraging the advantages of tendon-driven systems. Even though these robots have been intensively researched within the last decade, high-performance feedback control laws have only very recently been developed. In [1], a controller was developed utilizing Dynamic Surface Control (DSC), an extension to backstepping, with an adaptive neural network compensator for joint as well as muscle friction. We compare these novel control strategies to Computed Force Control (CFC), an existing technique from the field of tendondriven control, yielding highly improved trajectory tracking. The musculoskeletal robot Anthrob [2] serves as a benchmark.

Design and input saturation control with full-state constraints of lightweight tendon-driven musculoskeletal arm

“…Jantsch et al (2013) developed the simplified human upper limb with muscle tension sensors robot “ Anthrob ,” which comprised 13 compliant muscles and four degrees of freedom (DOF) joint. Kawaharazuka et al (2017a, 2017b) designed a human mimetic forearm with a radioulnar joint and completed routine tasks such as soldering, opening a book, turning a screw and swinging a badminton racket with low-precision requirements. Subsequently, they developed “ MusashiLarm ,” a complete musculoskeletal upper-limb platform comprising only joint modules, muscle modules, generic bone frames, muscle wire units and a few attachments (Kawaharazuka et al , 2019).…”

“…However, this algorithm needs to be trained thousands of times and cannot be applied to musculoskeletal robot platforms. Computational muscle control (Jantsch et al , 2012) and antagonist inhibition control (Kawaharazuka et al , 2017a, 2017b) are the most commonly used low-level control methods. Jantsch et al (2015) developed an adaptive neural network dynamic surface control to improve the trajectory tracking performance of “ Anthrob .” The results showed that the average angle error of the elbow joint was 0.12 rad.…”

Online Learning Feedback Control Considering Hysteresis for Musculoskeletal Structures

“…To avoid these problems, it is required to complete the feedback trial quickly or estimate its muscle Jacobian accurately. Although there are several methods estimating muscle Jacobian for musculoskeletal humanoids [10], [12], [13], there will always be some errors. Also, although there are methods to realize target posture using reinforcement learning [14], they are mostly performed in simulation only and are difficult to handle actual musculoskeletal humanoids with multiple degrees of freedom.…”

Bodily Aware Soft Robots: Integration of Proprioceptive and Exteroceptive Sensors

“…They are often underactuated [8], [12], which means that their control is not trivial. In the past few years data driven, model-free control methods have gained great interest due to their ability to learn from data and to approximate nonlinear, hyperelastic behaviour [13], [4]. This usually exploits artificial neural networks to create generalised models.…”

“…In the past few years data driven, model-free 1st Conference on Robot Learning (CoRL 2017), Mountain View, United States. arXiv:1710.05419v2 [cs.RO] 8 Nov 2017 control methods have gained great interest due to their ability to learn from data and to approximate nonlinear, hyperelastic behaviour [13], [4]. This usually exploits artificial neural networks to create generalised models.…”