Towards a more controllable sensorised soft gripper : a data-driven approach

Robotic grippers have been constantly improving over the years to become more dextrous and adaptable in handing difficult objects with variations in their shape or uncertainty in their positioning. A challenge that remains difficult until now is the ability to handle delicate objects that can be easily considered as defective due to their interaction with the gripper, such as the case for food products or finely machined parts. An interesting emerging approach to tackle this challenge is to rethink the origin of the problem, which is the fact that all conventional grippers are made of hard and rigid components that can easily damage objects during grasping if not precisely controlled based on reliable sensory feedback. Hence, creating gripper fingers from soft materials makes them inherently safe and relaxes the need for sophisticated sensing and complex control. However, several open research challenges exist that are hindering the full utilisation of soft robotic components. In the context of soft grippers, relying primarily on the soft nature of the fingers to passively and gently adapt to its targets although highly desirable, consequently means that no sensory feedback is available to have better control over the grasping process or confirm its success. In this research, a low-cost soft gripper was developed based on the ribbed pneumatic bending actuators with embedded bend sensing, in order to investigate the potential for sensor-guided control of soft gripper fingers. A purely data-driven approach is proposed that utilises basic sensory feedback to accurately estimate and control the bending of individual soft gripper fingers using simple empirical models that do not require any material characterisation or precise physical models. First, an experiment was designed to study the effect of varying the internal channel dimensions of soft finger samples with the same outer size, on their bending and force responses at variable input pressures. The results of this experiment provided useful design guidelines that can be followed to maximise the bending and force capabilities of the soft fingers and identified the best performing design of those tested. The experiment also illustrated how the soft finger's behaviour is governed by its designed morphology, which is consistent for fixed input conditions. The second step was to embed the soft fingers with resistive flex sensors, which change in resistance during bending without hindering the desired compliance. Additionally, onboard pressure sensors were used to measure the actual internal pressure developed inside the finger during actuation. Linear regression and artificial neural networks (ANN) are two common data-driven techniques that were implemented in this research. Both were fed with training data consisting of the flex and pressure measurements acquired by testing a soft finger sample at different pressure levels and orientations, with the corresponding synchronised bending angle measured using a vision system. The developed models were successfully validated using new data acquired at untrained conditions, with the ANN providing more accurate bending estimations at the expense of heavier computation. Lastly, a PID controller was developed which utilises the simple empirical model to estimate the current bending angle, calculates the error from a target value, and outputs a duty cycle value for the PWM signal regulating the supplied pressure. The controller was successful in controlling the modelled sensorised soft finger to accurately follow stepped and sinusoidal reference signals. Moreover, the combined multi-sensory feedback from the complete soft gripper was analysed to investigate the possibility of distinguishing between the free-bending and contact states, as well as differentiating between objects of different sizes. The main interest here was to evaluate if useful inferences can be made using the raw data from the flex and pressure sensors without having to model the real bending response of each soft finger individually. An experiment was conducted which involved grasping a set of objects of variable sizes and weights and collecting the resulting sensory feedback. The results of the experiment provided a clear relationship between the grasped object size and the averaged final flex sensor readings from opposing fingers supplied with the same pressure input. The results also showed the possibility of achieving contact detection by simply monitoring the current flex sensor's response during grasping and comparing it to the known free-bending response. A clear deviation can be witnessed at the occurrence of contact depending on the object size, which can be then used to stop the actuation. Furthermore, an interesting observation from this experiment was witnessed when monitoring the flex sensor's response during grasping against the measured internal pressure. Two distinct response curves were identified which reflects whether the object was grasped at the fingertips (precision) or encapsulated within the gripper (power), providing additional useful feedback about the grasp using simple sensory feedback. The last contribution of this research was the investigation of additive manufacturing as an alternative fabrication method to the manual multi-stage soft lithography technique. Automating the fabrication of soft grippers is not only desired for its speed and ease of use, but more importantly to improve the output consistency so that an empirical model derived for a specific actuator design can be potentially used for different samples with minimal need for updating. Functional soft finger based on the pleated morphology and flexible strain sensors were successfully 3D printed using a standard material extrusion-based printer after tuning the print parameters. The bending and force responses of the unit were experimentally characterised, and fatigue tests conducted to evaluate consistency. The printed soft finger was able to operate at higher pressures and hence generated larger contact forces while maintaining the desired compliance. Combining two of those units results in a two-fingered soft gripper that can be easily customised and directly printed in a single stage. The proposed data-driven modelling approach was successfully implemented using the printed finger as an additional validation to demonstrate the flexibility of using this approach with different actuator morphologies and materials. The outcomes of this investigation provided design guidelines and print settings recommended to successfully print air-tight soft fingers and highly flexible strain sensors. Finally, the results of this research deliver a simple purely data-driven approach for modelling and controlling soft grippers that are not limited to a specific morphology or material, as well as an automated process for fabricating those with better consistency. The key requirement is to generate relatively small datasets from simple, inexpensive sensors during the systematic experimental testing, as demonstrated in this research with the moulded and 3D printed soft gripper fingers. Ultimately, with innovations in additive manufacturing technologies enabling more difficult geometries and wider choices of flexible materials to be printed, combined with advanced machine learning algorithms processing larger grasping datasets, more dextrous sensorised soft grippers can be reliably printed to safely manipulate delicate targets in various real-life applications.