Controlling the softness
Soft robotics represents a paradigm shift in how machines interact with the physical world. Unlike rigid robots, soft robotic systems are built from compliant materials, allowing them to adapt safely and effectively to complex, changing environments. However, this flexibility comes at a cost: controlling such systems is notoriously difficult due to their highly non-linear and deformable dynamics. Backed by the Independent Research Fund Denmark, this research project addresses this challenge by developing advanced AI-based control methods inspired by the structure and function of the human brain.
Inspired by the brain
At the core of this project is the design of a three-level, brain-inspired control architecture that integrates reinforcement learning, classical control theory, and neuro-inspired computational models. The strength of this approach lies precisely in the combination: rather than relying on a single paradigm, the architecture brings together complementary capabilities that, when unified, enable a level of performance and robustness that none of the individual components could achieve alone.
- Reinforcement learning provides the system with the ability to learn from interaction. It enables the robot to discover control strategies through experience, adapting to complex environments and dynamics that are difficult to model explicitly. This is particularly important for soft robots, where precise analytical models are often unavailable. However, reinforcement learning alone can be unstable, data-intensive, and difficult to deploy safely in real-world systems.
- Control theory addresses these limitations by introducing structure, stability, and guarantees on system behavior. It provides well-established tools to ensure that the robot behaves predictably and remains within safe operating conditions. Yet, traditional control methods struggle when faced with the high-dimensional, non-linear, and uncertain dynamics of soft materials.
- Neuro-inspired models add a third dimension by introducing principles observed in biological systems, particularly in terms of efficiency and adaptability. These models are designed to process information in a distributed, event-driven manner, enabling fast and energy-efficient responses. They also support more natural forms of learning and adaptation, reflecting how biological organisms control movement in uncertain environments.
By combining these three approaches, the architecture achieves a balance between adaptability, stability, and efficiency. Reinforcement learning drives exploration and reward mechanisms control theory ensures reliability, and neuro-inspired models enable scalable and energy-efficient computation. This layered integration mirrors the organization of the human brain, where reflexive control, learned behavior, and higher-level reasoning coexist and interact.
The project strengthens this architecture by integrating advanced AI techniques. Spiking neural networks (SNNs), inspired by biological neurons, process information through discrete spikes, making them well-suited to temporal processing and far more energy-efficient than traditional neural networks – an important advantage for real-world robotic deployment.
Distributional reinforcement learning models the full range of possible outcomes rather than just the expected result. This helps the system capture uncertainty and variability, enabling more robust decision-making in complex environments.
Chain-of-thought reasoning adds structured, interpretable decision-making by allowing the system to work through intermediate steps before acting – effectively “thinking” before acting This improves transparency while supporting more complex, goal-directed behaviour.
Together, these components create a hierarchical control system that operates across different time scales and levels of abstraction, from fast reactions to slower, deliberative reasoning.
Real-world applications
Soft robotics has enormous potential in human-centered domains. In healthcare, these systems can assist in minimally invasive surgery, rehabilitation, and targeted drug delivery. Their inherent safety and compliance make them ideal for working alongside humans, particularly in eldercare and assistive technologies.
The development of energy-efficient and adaptive control systems is critical for deploying such robots outside laboratory settings. This project directly contributes to addressing major societal challenges, including aging populations, the need for sustainable technologies, and the automation of complex tasks in sensitive environments. By enabling safer and more intelligent human/robot interaction, the research supports a future where robots seamlessly integrate into daily life.
