Soft and Neuromorphic Robotics—Enhancing Soft Robotic Grippers Through In-Hardware Trainable Behaviour

To ensure proper and safe functioning, robots need to interact with their immediate surroundings adaptively. While novel methods to incorporate adaptive behaviour are currently being researched, they can be placed into two categories: passive and active adaptivity. Passive adaptivity exploits flexibility in robotic hardware design, allowing contact energy to be absorbed through deformation, resulting in robust interaction when interacting with the world without increasing the computational load. Meanwhile, active adaptivity relies on data processing to analyse and respond to stimuli, providing additional safety measures and control options through active analysis of the environment. Within this context, we focus on the benefits adaptivity provides when handling and manipulating objects with a robotic gripper. For passive adaptivity, we use a gripper with soft actuators enhanced with rigid components to obtain a strong, stable, and shape-conforming grasp. This hybrid soft-rigid design inspired by the human hand provides a basis to incorporate active adaptive properties.

When introducing further adaptive behaviour into a robotic gripper, the human body provides a lot of inspiration for adaptive properties that can be incorporated. One intuitive example is the heat avoidance reflex, which we cannot incorporate using the flexible elements of our soft-rigid gripper, requiring us to use a form of active adaptivity to incorporate this behaviour into our system. For this, we turn to neuromorphic computing, which mimics neuronal and synaptic behaviour through analog circuits using organic electrochemical transistors (OECTs). This allows us to integrate learnable adaptive behaviour while also allowing for lower power requirements than for conventional Artificial Neural Networks (ANNs). To incorporate reflexive behaviour and provide our systems with the capacity to learn from previous mistakes, we design small, function-driven circuits that can teach our robot to respond to one stimulus at a time, such as the heat avoidance reflex. These circuits provide a measurable output voltage that reacts to environmental changes and adjusts its sensitivity to avoid undesired scenarios, allowing our system to respond to complex scenarios.

We show through a dynamic simulation that the circuit can allow a mass to avoid high temperatures, where repeated exposure improves the sensitivity, making the high-temperature avoidance response stronger. Expanding on this principle, we aim to simulate a robotic system in more detail while focussing on OECT dynamics and exploring responses to additional stimuli. Following simulations, we plan to build physical circuits and apply them to the soft-rigid robot to validate all simulated findings. Each step brings us closer to a combined soft-rigid and neuromorphic gripper that is intrinsically adaptive, strong, and trainable while maintaining low power requirements.