The control of multi-legged animal walking is a neuromechanical process where the nervous system produces neural signals to activate the muscles for driving the skeletons (e.g., legs and joints). To model and achieve this process on an artificial multi-legged machine is a difficult and challenging problem. This is because the process needs to model the interaction among the redundant nervous system (i.e., multiple neurons and synapses) and the redundant musculoskeletal systems (i.e., dozens of muscles and joints). Moreover, the two (i.e., neural and musculoskeletal) intrinsic redundancies lead to the kinematic redundancy where determining the joint angles that result in the particular positions at the endpoints of the legs is a tough task. The neurophysiologist Nikolai Bernstein pointed out, these three redundancies cause the difficulties of modeling and achieving a neuromechanical process on an artificial system with many degrees of freedom (DOFs). The modeled neuromechanical process needs in real time 1) to coordinate very many DOFs of jointed legs and muscle-like mechanisms, 2) to generate the proper leg stiffness (i.e., compliance), and 3) to determine joint angles that give rise to particular positions at the endpoints of the legs. To show a way forward to model these three sub-processes, here we develop and implement neuromechanical controller coupled with sensorimotor learning on the hexapod robot AMOS with 19 DOFs. The controller consists of the modular neural network (MNN) for coordinating 18 joints and several virtual agonist-antagonist muscle mechanisms (VAAMs) for variable compliant joint motions. In addition, sensorimotor learning, including forward models and dual-rate learning processes, is introduced for predicting foot force feedback and for online adjusting the VAAMs’ stiffness parameters. The control and learning mechanisms enable the hexapod robot AMOS to achieve adaptive compliant walking that adapts to different gaits (e.g., tripod gaits) and rough surfaces (e.g., gravel). As a result, AMOS can not only well classify rough surfaces, but also perform more energy-efficient walking on them, compared to other small legged robots. In addition, the developed method shows that the tight combination of neural control with tunable muscle-like functions, guided by sensory feedback and coupled with sensorimotor learning, is a way forward to better understand and solve adaptive coordination problems in multi-legged locomotion. In this thesis, the presented work has seven key contributions: 1) it develops a simplified muscle model for the muscle-like functions that underlie variable compliant joint motions; 2) it presents the underlying muscle-like functions (e.g., brakes) of the VAAMs that are comparable to those of biological muscles; 3) it utilizes a proximodistal gradient of neural control and muscle functions to enhance the stability of legged robot locomotion under adaptive compliance control; 4) it achieves variable compliant joint motions relying only on force sensing at the end effectors of the legs. Thus, the implementation reduces sensing and design efforts of legged robots; 5) it exploits the compliant joint signals generated by the VAAMs to well classify surfaces without using multiple sensing; 6) it utilizes sensorimotor learning to self-adjust the stiffness parameters of the VAAMs that adapts to insect-like gaits (e.g., tripod gaits) and surfaces (e.g., gravel), thereby leading to more energy-efficient hexapedal walking; 7) it provides a way forward to model stable and adaptive compliant insect-like walking, by implementing a bio-inspired neural network and several virtual agonist-antagonist mechanisms (VAAMs) on the hexapod robot AMOS with 19 DOFs (i.e., probing Bernstein’s ‘degrees of freedom’ problem).
|Date of defence||8. Jun 2015|
|Publication status||Published - 12. Jun 2015|