Visuo-dynamic self-modelling of soft robotic systems.

Soft robots exhibit complex nonlinear dynamics with large degrees of freedom, making their modelling and control challenging. Typically, reduced-order models in time or space are used in addressing these challenges, but the resulting simplification limits soft robot control accuracy and restricts their range of motion. In this work, we introduce an end-to-end learning-based approach for fully dynamic modelling of any general robotic system that does not rely on predefined structures, learning dynamic models of the robot directly in the visual space. The generated models possess identical dimensionality to the observation space, resulting in models whose complexity is determined by the sensory system without explicitly decomposing the problem. To validate the effectiveness of our proposed method, we apply it to a fully soft robotic manipulator, and we demonstrate its applicability in controller development through an open-loop optimization-based controller. We achieve a wide range of dynamic control tasks including shape control, trajectory tracking and obstacle avoidance using a model derived from just 90 min of real-world data. Our work thus far provides the most comprehensive strategy for controlling a general soft robotic system, without constraints on the shape, properties, or dimensionality of the system.

3D Printable Soft Sensory Fiber Networks for Robust and Complex Tactile Sensing.

The human tactile system is composed of multi-functional mechanoreceptors distributed in an optimized manner. Having the ability to design and optimize multi-modal soft sensory systems can further enhance the capabilities of current soft robotic systems. This work presents a complete framework for the fabrication of soft sensory fiber networks for contact localization, using pellet-based 3D printing of piezoresistive elastomers to manufacture flexible sensory networks with precise and repeatable performances. Given a desirable soft sensor property, our methodology can design and fabricate optimized sensor morphologies without human intervention. Extensive simulation and experimental studies are performed on two printed networks, comparing a baseline network to one optimized via an existing information theory based approach. Machine learning is used for contact localization based on the sensor responses. The sensor responses match simulations with tunable performances and good localization accuracy, even in the presence of damage and nonlinear material properties. The potential of the networks to function as capacitive sensors is also demonstrated.

Closing the Control Loop with Time-Variant Embedded Soft Sensors and Recurrent Neural Networks.

Embedded soft sensors can significantly impact the design and control of soft-bodied robots. Although there have been considerable advances in technology behind these novel sensing materials, their application in real-world tasks, especially in closed-loop control tasks, has been severely limited. This is mainly because of the challenge involved with modeling a nonlinear time-variant sensor embedded in a complex soft-bodied system. This article presents a learning-based approach for closed-loop force control with embedded soft sensors and recurrent neural networks (RNNs). We present learning protocols for training a class of RNNs called long short-term memory (LSTM) that allows us to develop accurate and robust state estimation models of these complex dynamical systems within a short period of time. Using this model, we develop a simple feedback force controller for a soft anthropomorphic finger even with significant drift and hysteresis in our feedback signal. Simulation and experimental studies are conducted to analyze the capabilities and generalizability of the control architecture. Experimentally, we are able to develop a closed-loop controller with a control frequency of 25 Hz and an average accuracy of 0.17 N. Our results indicate that current soft sensing technologies can already be used in real-world applications with the aid of machine learning techniques and an appropriate training methodology.

Manipulation of free-floating objects using Faraday flows and deep reinforcement learning.

The ability to remotely control a free-floating object through surface flows on a fluid medium can facilitate numerous applications. Current studies on this problem have been limited to uni-directional motion control due to the challenging nature of the control problem. Analytical modelling of the object dynamics is difficult due to the high-dimensionality and mixing of the surface flows while the control problem is hard due to the nonlinear slow dynamics of the fluid medium, underactuation, and chaotic regions. This study presents a methodology for manipulation of free-floating objects using large-scale physical experimentation and recent advances in deep reinforcement learning. We demonstrate our methodology through the open-loop control of a free-floating object in water using a robotic arm. Our learned control policy is relatively quick to obtain, highly data efficient, and easily scalable to a higher-dimensional parameter space and/or experimental scenarios. Our results show the potential of data-driven approaches for solving and analyzing highly complex nonlinear control problems.

Soft Self-Healing Fluidic Tactile Sensors with Damage Detection and Localization Abilities.

Self-healing sensors have the potential to increase the lifespan of existing sensing technologies, especially in soft robotic and wearable applications. Furthermore, they could bestow additional functionality to the sensing system because of their self-healing ability. This paper presents the design for a self-healing sensor that can be used for damage detection and localization in a continuous manner. The soft sensor can recover full functionality almost instantaneously at room temperature, making the healing process fully autonomous. The working principle of the sensor is based on the measurement of air pressure inside enclosed chambers, making the fabrication and the modeling of the sensors easy. We characterize the force sensing abilities of the proposed sensor and perform damage detection and localization over a one-dimensional and two-dimensional surface using multilateration techniques. The proposed solution is highly scalable, easy-to-build, cheap and even applicable for multi-damage detection.

Learning to stop: a unifying principle for legged locomotion in varying environments.

Evolutionary studies have unequivocally proven the transition of living organisms from water to land. Consequently, it can be deduced that locomotion strategies must have evolved from one environment to the other. However, the mechanism by which this transition happened and its implications on bio-mechanical studies and robotics research have not been explored in detail. This paper presents a unifying control strategy for locomotion in varying environments based on the principle of 'learning to stop'. Using a common reinforcement learning framework, deep deterministic policy gradient, we show that our proposed learning strategy facilitates a fast and safe methodology for transferring learned controllers from the facile water environment to the harsh land environment. Our results not only propose a plausible mechanism for safe and quick transition of locomotion strategies from a water to land environment but also provide a novel alternative for safer and faster training of robots.

First-Order Dynamic Modeling and Control of Soft Robots.

Modeling of soft robots is typically performed at the static level or at a second-order fully dynamic level. Controllers developed upon these models have several advantages and disadvantages. Static controllers, based on the kinematic relations tend to be the easiest to develop, but by sacrificing accuracy, efficiency and the natural dynamics. Controllers developed using second-order dynamic models tend to be computationally expensive, but allow optimal control. Here we propose that the dynamic model of a soft robot can be reduced to first-order dynamical equation owing to their high damping and low inertial properties, as typically observed in nature, with minimal loss in accuracy. This paper investigates the validity of this assumption and the advantages it provides to the modeling and control of soft robots. Our results demonstrate that this model approximation is a powerful tool for developing closed-loop task-space dynamic controllers for soft robots by simplifying the planning and sensory feedback process with minimal effects on the controller accuracy.

Control Strategies for Soft Robotic Manipulators: A Survey.

With the rise of soft robotics technology and applications, there have been increasing interests in the development of controllers appropriate for their particular design. Being fundamentally different from traditional rigid robots, there is still not a unified framework for the design, analysis, and control of these high-dimensional robots. This review article attempts to provide an insight into various controllers developed for continuum/soft robots as a guideline for future applications in the soft robotics field. A comprehensive assessment of various control strategies and an insight into the future areas of research in this field are presented.

Modeling the Encoding of Saccade Kinematic Metrics in the Purkinje Cell Layer of the Cerebellar Vermis.

Recent electrophysiological observations related to saccadic eye movements in rhesus monkeys, suggest a prediction of the sensory consequences of movement in the Purkinje cell layer of the cerebellar oculomotor vermis (OMV). A definite encoding of real-time motion of the eye has been observed in simple-spike responses of the combined burst-pause Purkinje cell populations, organized based upon their complex-spike directional tuning. However, the underlying control mechanisms that could lead to such action encoding are still unclear. We propose a saccade control model, with emphasis on the structure of the OMV and its interaction with the extra-cerebellar components. In the simulated bilateral organization of the OMV, each caudal fastigial nucleus is arranged to receive incoming projections from combined burst-pause Purkinje cell populations. The OMV, through the caudal fastigial nuclei, interacts with the brainstem to provide adaptive saccade gain corrections that minimize the visual error in reaching a given target location. The simulation results corroborate the experimental Purkinje cell population activity patterns and their relation with saccade kinematic metrics. The Purkinje layer activity that emerges from the proposed organization, precisely predicted the speed of the eye at different target eccentricities. Simulated granular layer activity suggests no separate dynamics with respect to shaping the bilateral Purkine layer activity. We further examine the validity of the simulated OMV in maintaining the accuracy of saccadic eye movements in the presence of signal dependent variabilities, that can occur in extra-cerebellar pathways.

Learning Closed Loop Kinematic Controllers for Continuum Manipulators in Unstructured Environments.

This article introduces a machine-learning-based approach for closed loop kinematic control of continuum manipulators in the task space. For this purpose, we propose a unique formulation for learning the inverse kinematics of a continuum manipulator while integrating end-effector feedback. We demonstrate that this model-free approach for kinematic control is very well suited for nonlinear stochastic continuum robots. The article addresses problems that are vital for practical realization of machine-learning techniques. The primary objective is to solve the redundancy problem while making the algorithm scalable, fast, and tolerant to stochasticity, requiring minimal sensor elements and involving few open parameters for tuning. In addition, we demonstrate that the proposed controller can exhibit adaptive behavior in the presence of external forces and in an unstructured environment with the help of the morphological properties of the manipulator. Experimental validation of the proposed controller is done on a six-degree-of-freedom tendon-driven manipulator for pose control of the end effector in three-dimensional space with and without external forces. The experimental results exhibit accurate, reliable, and adaptive behavior of the proposed system, which appears suitable for the field of continuum service robots.