Advancing Kinematic Control from Rigid Robots to Dynamic Bio-Inspired Systems for Adaptive Behaviours

Despite the increasing global population, widespread staffing shortages continue due to difficulties in recruitment, staff retention, and rising labour costs. There is, therefore, an urgent need to automate tasks beyond traditional factory settings through robotics. However, robots and their designers must overcome the harsh, unpredictable, and unstructured conditions encountered from development to deployment. One promising approach to this challenge is embodied intelligence, where system behaviour emerges from the interaction between the brain, body, and environment. Yet, the embodiment of commercially available robots differs significantly from that of biological organisms. Most commercial robots have rigid structures and actuation, whereas animals possess materials with varying stiffness, such as skin, tissue, muscles, and bones, resulting in rich dynamics that expand behavioural capabilities. However, these bio-inspired systems are far more difficult to control due to complex kinematics and dynamics. While various control strategies exist, closed-loop kinematic control offers the advantage of implicitly managing quasi-static dynamic internal and external forces. This thesis presents a conceptual framework for harnessing the embodied dynamics of bio-inspired systems through developmentally inspired sensory-motor coordination. This approach enables the learning of sensory-motor correlations and facilitates kinematic control with emergent flexible behaviours. However, deploying artificial systems in real-world settings requires consideration of additional factors beyond adaptability, including cost, task performance, and technological availability, all of which are determined by the task being automated. In pursuit of developing more robust robots for real-world applications, this thesis examines critical aspects of kinematic control systems for both rigid and bio-inspired morphologies, spanning from computational intelligence through machine learning to embodied intelligence in bio-inspired systems. Several novel contributions are made. The first two are industrial case studies addressing previously unautomated tasks: potato planting and road-side litter collection, with PotatoBot and LitterBot, respectively. The limitations of rigid systems with kinematic control were mitigated by integrating modular, agnostic components (with one another), allowing for improvability, reusability, and task-specific specialisation. Combining agnostic modular components with laboratory replication of key real-world conditions and an agile-like development approach, proved essential for avoiding the logistical and safety challenges of field testing. This strategy facilitated high initial task performance while minimising the additional hardware integration, development time, and task-specialisation costs associated with adapting rigid, off-the-shelf cobot arms for real-world tasks. The third contribution demonstrates the benefits of sensory-motor coordination in bio-inspired systems for another labour-intensive task: identifying fruit ripeness. This led to the development of a novel Electrical Impedance Tomography (EIT)-enabled soft gripper. Its adaptive morphology and interaction with the environment structured the high-dimensional, 'through-the-material' sensing, enabling the mapping of this data to low-dimensional ripeness properties for accurate predictions. The final three contributions directly apply the proposed framework to dynamic bio-inspired systems, including simulated musculoskeletal models and soft continuum robots. Using developmentally inspired stochastic data-sampling methods, such as Motor Babbling with feedforward neural networks, and Spontaneous Muscle Activations with Hebbian-based learning, the systems' dynamics induced statistical regularities between sensory and motor systems. This facilitated the learning of pseudo-inverse Jacobian approximations during the learning phase, which were later implemented in closed-loop feedback control during deployment. This framework eliminates the need for separate path planning, complete prior system knowledge, or analytical modelling of kinematics and dynamics. It also leverages motor redundancy to adapt to challenges such as unknown payloads, obstructions, and actuator failures through the interaction between the controller, body, and environment. The final chapter discusses the trade-offs between rigid and bio-inspired embodiments and compares the two learning-based kinematic control methods for bio-inspired systems. It also explores future research directions to build upon the work presented in this thesis.

Name dc:type.qualificationname

Doctor of Philosophy (PhD)

Level dc:type.qualificationlevel

Doctoral

Grantor dc:publisher.institution

University of Cambridge

Year dc:date.issued

2025

Author dc:creator

• Almanzor, Elijah

Advisor dc:contributor.advisor

• Iida, Fumiya

• Bio-inspired Robotics
• Learning-based Control
• Soft Robots
• Musculoskeletal Systems
• Industrial Robotics

Licence

• https://www.repository.cam.ac.uk/bitstreams/dc749309-d996-4687-92d9-a4de73acac0c/download
• http://purl.org/NET/rdflicense/allrightsreserved

Language dc:language

eng

Author Identifier

0000-0001-8101-4217

OAI identifier oai:identifier

oai:www.repository.cam.ac.uk:1810/389057