Intelligent Control of Soft Snake Robot Locomotion with Biomimic Vertebrate System

Liu, Xuan

Etd

Intelligent Control of Soft Snake Robot Locomotion with Biomimic Vertebrate System

Public Deposited

Soft snake robots have unique advantages in traversing through cluttered and confined environments because they are equipped with highly flexible body structures, deformable materials, and high sensitivity almost in any part of their body. These advantages have resulted in great expectations for the snake robots in many difficult applications, including social rescuing, cave exploration and medical operations, etc. However, planning and control of such types of robots remains a challenging problem, as these robots have infinitely many degrees of freedom (DoFs) in their body links, and soft actuators with hard-to-identify dynamics. By taking inspiration from the cerebral and spinal control of rhythmic behaviors in natural animals, we in this thesis develop bio-inspired locomotion controllers that allow the robot to freely sense and explore the environment, and flexibly maneuver its locomotion modes with embodied intelligence in a planar workspace. These controllers are based on the concept of Central Pattern Generators (CPGs), which are a series of mathematical models that describe spinal neural circuits' activities that generate the rhythmic patterns for animals' organ contraction and locomotion. Among the popular CPGs in bio-inspired robot control, the Matsuoka oscillator is well-known for generating high-fidelity rhythmic neural oscillation patterns for robot locomotion gaits in mimicking animal behaviors. However, its nonlinearity makes it harder to analyze the system state properties and gradually becomes less popular in robotics locomotion studies. During my study on the learning-based locomotion of soft robot snakes, we found that the rhythmic patterns in the Matsuoka oscillator can be easily learned and maneuvered by a model-free RL algorithm. On the basis of Matsuoka's theory, we further justified that the oscillation features of the Matsuoka oscillator, including oscillation bias, frequency, and amplitude have a clear relation with the specific coefficients of the Matsuoka oscillator, and therefore can be efficiently controlled by the RL agent. Such a mechanism allows the proposed control framework to easily learn flexible steering and speed control of the soft snake robot when tracking dynamically changing goals. However, as we tried to incorporate the sensory feedback mechanism in the Matsuoka CPG system to realize the contact-aware locomotion of our soft snake robot, we encountered a problem such that the conventional feedback approach could bring significant overshoot and delay to the oscillation patterns of the Matsuoka oscillator, and therefore impede the performance of the whole RL-CPG control scheme during the contact-aware locomotion of the robot. To solve this issue, we develop a novel sensory feedback mechanism for the Matsuoka CPG network. This mechanism allows the Matsuoka CPG system to work like a ``spine cord'' in the whole contact-aware control scheme, which simultaneously takes the stimuli including tonic input signals from the ``brain'' (a goal-tracking locomotion controller) and sensory feedback signals from the ``reflex arc'' (the contact reactive controller), and generates rhythmic signals to actuate the soft snake robot to slither through densely allocated obstacles. In the design of the ``reflex arc'', we develop two distinctive types of reactive controllers -- 1) a reinforcement learning (RL) sensor regulator that learns to manipulate the sensory feedback inputs of the CPG system, and 2) a local reflexive sensor-CPG network that directly connects sensor readings and the CPG's feedback inputs in a specific topology. These two reactive controllers respectively facilitate two different contact-aware locomotion control schemes. In summary, the original contribution of this thesis can be organized in two folds: 1. In theory, we have first analyzed and proved the Matsuoka CPG's steering maneuverability to allow an organic composition of the RL module and CPG module to form an efficient learning-based locomotion controller, which is also tested to be generalizable to other robotic platforms. In addition, we have developed free-response oscillation constraints (FOC) of the Matsuoka CPG system for sim-to-real transfer. Last but not least, we have completed the development of the sensory feedback mechanism in Matsuoka CPG system. Such a mechanism is combined with two feedback reactive controllers based on two different theories (hybrid control and local reflexive controller) to realize contact-aware locomotion of the soft snake robot. 2. In practice, we design and build three generations of soft snake robots to optimize their performance and expand their functionality. The optimality and robustness of the proposed control design are validated in both simulated and real soft snake robots, along with a sufficient comparison to other methods (including other RL and conventional Matsuoka CPG systems). The contact-aware locomotion control schemes are tested and evaluated in both simulated and real soft snake robots, showing promising performance in the contact-aware locomotion tasks. Our experimental results have validated the advantages of the Matsuoka CPG system for bio-inspired robot controller design. Overall, our series work is based on the theory and application of the Matsuoka oscillator, including the discussion and derivation of the special properties of the CPG system. The contribution covers hardware design and manufacturing, control scheme design (including locomotion control and sensory feedback control), and experiment design and implementation (including simulation and reality). It makes a significant breakthrough in the research of bio-inspired robot control.

Creator