When a human controls a robot directly or via teleoperation, incomplete information and unpredictable environmental conditions can lead to conflicts between their plans. Differential game theory (GT) offers a framework for optimal robotic assistance, but existing methods for identifying the human model require a shared plan. This article introduces an approach to deal with a diverging human plan by leveraging a neuromechanical model of their viscoelasticity to directly estimate their motion intent during movement. The viscoelastic gains can then be integrated into a GT framework to compute optimal contribution of the robot to the common motor task. We evaluated the proposed method in experiments, comparing it with fixed impedance control and nonoptimal variable impedance control. Results demonstrate stable interactions even in the presence of conflicting motion plans, and superior performance relative to these two alternative methods.

Due to high Degrees of Freedom (DoF), dexterous hand is capable of manipulating tools with heavily complex, uncertain and unstructured industrial environments, thereby enabling stronger coupling with surroundings and human colleagues. However, the high-dimensional state and multi-point sliding contacts pose significant challenges in modeling and policy transfer within different domain during Hand-Tool-Environment (HTE) interactions. In this paper, a progressive generative representation (PGR), mapping observation into low-dim latent space, is employed to extract contact-rich feature and facilitate adaptive transfer under domain shift with its inherent randomness and progressive adjustment. Furthermore, a model-based optimal control is effectively integrasted with the implicit latent dynamics via temporal difference learning to infer manipulation policy. The proposed method is validated in a simulated industrial scenario involving ultrasound inspection task with ablation study, demonstrating both its effectiveness and necessity through point tracking and the similarity to predefined trajectory. Proposed method demonstrated superior performance in domain transfer involving stiffness and friction alterations when compared to all control groups. Notably, performance was restored to source domain levels with the application of merely 10 rollouts updates.

Neural architecture search (NAS) has shown great promise in automatically designing neural network models. Recently, block-wise NAS has been proposed to alleviate deep coupling problem between architectures and weights existed in the well-known weight-sharing NAS, by training the huge weight-sharing supernet block-wisely. However, the existing block-wise NAS methods, which resort to either supervised distillation or self-supervised contrastive learning scheme to enable block-wise optimization, take massive computational cost. To be specific, the former introduces an external high-capacity teacher model, while the latter involves supernet-scale momentum model and requires a long training schedule. Considering this, in this work, we propose a resource-friendly deeply supervised block-wise NAS (DBNAS) method. In the proposed DBNAS, we construct a lightweight deeply-supervised module after each block to enable a simple supervised learning scheme and leverage ground-truth labels to indirectly supervise optimization of each block progressively. Besides, the deeply-supervised module is specifically designed as structural and functional condensation of the supernet, which establishes global awareness for progressive block-wise optimization and helps search for promising architectures. Experimental results show that the DBNAS method only takes less than 1 GPU day to search out promising architectures on the ImageNet dataset with less GPU memory footprint than the other block-wise NAS works. The best-performing model among the searched DBNAS family achieves 75.6% Top-1 accuracy on ImageNet, which is competitive with the state-of-the-art NAS models. Moreover, our DBNAS family models also achieve good transfer performance on CIFAR-10/100, as well as two downstream tasks: object detection and semantic segmentation.

A key challenge in Meta-Reinforcement Learning (meta-RL) is the task distribution shift, since the generalization ability of most current meta-RL methods is limited to tasks sampled from the training distribution. In this paper, we propose Posterior Sampling Bayesian Lifelong In-Context Reinforcement Learning (PSBL), which is robust to task distribution shift. PSBL meta-trains a variant of transformer to directly perform amortized inference about the Predictive Posterior Distribution (PPD) of the optimal policy. Once trained, the network can infer the PPD online with frozen parameters. The agent then samples actions from the approximate PPD to perform online exploration, which progressively reduces uncertainty and enhances performance in the interaction with the environment. This property is known as in-context learning. Experimental results demonstrate that PSBL significantly outperforms standard Meta RL methods both in tasks with sparse rewards and dense rewards when the test task distribution is strictly shifted from the training distribution.

Aerial-aquatic vehicles are capable to move in the two most dominant fluids, making them more promising for a wide range of applications. We propose a prototype with special designs for propulsion and thruster configuration to cope with the vast differences in the fluid properties of water and air. For propulsion, the operating range is switched for the different mediums by the dual-speed propulsion unit, providing sufficient thrust and also ensuring output efficiency. For thruster configuration, thrust vectoring is realized by the rotation of the propulsion unit around the mount arm, thus enhancing the underwater maneuverability. This paper presents a quadrotor prototype of this concept and the design details and realization in practice.

Autonomous driving is currently one of the most popular topics, and computer vision is the key and perception-related technology in this field. Although images have the ability to contain a large amount of crucial feature information, deep learning methods require a significant number of samples, resulting in low data utilization. This paper proposes a learning from demonstration approach that leverages the control laws from human expert drivers to guide a deep convolutional neural network (CNN) in learning features from RGB images captured by the forward-facing camera. Subsequently, the CNN is utilized as a differentiable control layer to obtain an autonomous driving agent with the capabilities of an expert driver. Experimental results of this method are analyzed for designing a lane-keeping control system that learns human driving style via behavioral cloning.

A crucial aspect of human-robot integration is the implementation of person-following robots. However, autonomous robots continue to face challenges in tracking individuals within complex environments with dynamic obstacles. Traditional approaches to combining global and local planning are inadequate for this task due to the high level of uncertainty in the environment and the flexible behavior of the following target. To address these challenges, this paper proposes a framework for motion planning for wheeled robots. This method integrates dynamic obstacle avoidance and dynamic target following into an optimization problem based on model predictive control (MPC) with terminal constraint set to guarantee the safety of the following task in dynamic obstacle environments. Therefore, in this paper, a high-frequency state estimator is used to predict human behavior, and a spring model is used to model dynamic obstacles in the environment to keep the robot away from the obstacles. The effect of the person following is thoroughly tested in simulation with multiple scenarios and the comparison experiment, verifying the real-time effectiveness of the method.

Data-driven models have been revealed to be vulnerable to adversarial examples. However, the difficulty of data collection for different fault types in the actual industrial process is different, and the resulting imbalanced data problem brings greater difficulties to adversarial robustness learning, and brings the negative impact of poor robustness and majority class overfitting to common empirical adversarial training methods. To address these issues, we design a smoothing robust learning strategy. With linearly augmented adversarial data, balance calibration is imposed by class-aware label smoothing, and then the data related soft-label regularization is merged. Simultaneously, the imbalance of adversarial variants generation within the training is penalized to achieve a better adversarial game relationship. Case studies on industrial benchmarks Tennessee Eastman process (TEP) and NEU surface defect database (SDD) demonstrate that our approach achieves better standard and robust generalization.

Deep convolutional neural networks (DNNs) have been widely used in many applications, particularly in machine vision. It is challenging to accelerate DNNs on embedded systems because real-world machine vision applications should reserve a lot of external memory bandwidth for other tasks, such as video capture and display, while leaving little bandwidth for accelerating DNNs. In order to solve this issue, in this study, we propose a high-throughput accelerator, called reconfigurable tiny neural network accelerator (ReTiNNA), for the bandwidth-limited system and present a real-time object detection system for the high-resolution video image. We first present a dedicated computation engine that takes different data mapping methods for various filter types to improve data reuse and reduce hardware resources. We then propose an adaptive layer-wise tiling strategy that tiles the feature maps into strips to reduce the control complexity of data transmission dramatically and to improve the efficiency of data transmission. Finally, a design space exploration (DSE) approach is presented to explore design space more accurately in the case of insufficient bandwidth to improve the performance of the low-bandwidth accelerator. With a low bandwidth of 2.23 GB/s and a low hardware consumption of 90.261K LUTs and 448 DSPs, ReTiNNA can still achieve a high performance of 155.86 GOPS on VGG16 and 68.20 GOPS on ResNet50, which is better than other state-of-the-art designs implemented on FPGA devices. Furthermore, the real-time object detection system can achieve a high object detection speed of 19 fps for high-resolution video.

Manipulation in a densely cluttered environment creates complex challenges in perception to close the control loop, many of which are due to the sophisticated physical interaction between the environment and the manipulator. Drawing from biological sensory-motor control, to handle the task in such a scenario, tactile sensing can be used to provide an additional dimension of the rich contact information from the interaction for decision making and action selection to manoeuvre towards a target. In this paper, a new tactile-based motion planning and control framework based on bioinspiration is proposed and developed for a robot manipulator to manoeuvre in a cluttered environment. An iterative two-stage machine learning approach is used in this framework: an autoencoder is used to extract important cues from tactile sensory readings while a reinforcement learning technique is used to generate optimal motion sequence to efficiently reach the given target. The framework is implemented on a KUKA LBR iiwa robot mounted with a SynTouch BioTac tactile sensor and tested with real-life experiments. The results show that the system is able to move the end-effector through the cluttered environment to reach the target effectively.

This paper works on multiple-pursuer single-evader (MPSE) problems with a fast evader, which means multiple pursuers try to capture one evader while the evader tries to escape from the encirclement. The biggest concern is that the maximum velocity of the evader is larger than all the pursuers. Some improved strategies for the evader and pursuers based on traditional algorithms are firstly provided. Then gene expression programming (GEP) is used to generate new strategies which are better than the traditional ones. This paper shows configurations of function set, terminal set, fitness, evaluation function, and other parameters used in the GEP method, which can be implemented in other cases or similar problems.

Robotic systems are increasingly required not only to generate precise motions to complete their tasks but also to handle the interactions with the environment or human. Significantly, soft interaction brings great challenges on the force control due to the nonlinear, viscoelastic and inhomogeneous properties of the soft environment. In this paper, a robust impedance control scheme utilizing integral backstepping technology and integral terminal sliding mode control is proposed to achieve force tracking for an ultrasonic motor-actuated end-effector in a soft environment. In particular, the steady-state performance of the target impedance while in contact with soft environment is derived and analyzed with the nonlinear Hunt-Crossley model. Finally, the dynamic force tracking performance of the proposed control scheme is verified via several experiments.

A vision-based cooperative formation control method is proposed in this paper for a heterogeneous unmanned system including an UAV (Unmanned Aerial Vehicle) and multiple UGVs (Unmanned Ground Vehicles). Considering the supervisory role of the UAV and the time-varying relative localization between UAV and UGVs, we aim at controlling the multi-UGVs to a desired formation relying only on the visual information obtained by a camera mounted on the UAV. Meanwhile, the UGV group is driven to track the flying UAV using a feedback control algorithm. A gradient descent-like control scheme which considers the visual sensing range constraint of the camera is thus adopted based on a designed cost function. Finally, the proposed method has been successfully validated through simulations.

We propose a biomimetic approach for steering motion control of a snake robot. Inspired by a vertebrate biological motor system paradigm, a hierarchical control scheme is adopted. In the control scheme, an artificial central pattern generator (CPG) is employed to generate serpentine locomotion in the robot. This generator outputs the coordinated desired joint angle commands to each lower-level effector controller, while the locomotion can be controlled through CPG modulation by a higher-level motion controller. The motion controller consists of a cerebellar model articulation controller (CMAC) and a proportional-derivative (PD) controller. Because of the fast learning ability of the CMAC, the proposed motion controller can drive the robot to track the desired orientation and adapt to unexpected perturbations. The PD controller is employed to expedite the convergence speed of the motion controller. Finally, both numerical studies and experiments proved that the proposed approach can help the snake robot achieve good tracking performance and adaptability in a varying environment.

Soft robots have recently evoked extensive attention due to their abilities to work effectively in unstructured environments. As an actuation technology of soft robots, dielectric elastomers exhibit many intriguing attributes such as large strain and high energy density. This work presents a novel dielectric elastomer based soft crawling robot inspired by inchworms. To fill the need of control of the soft robot, a model describing the interaction between the dielectric elastomer actuator and the environment is proposed, which takes inertia, viscoelasticity and friction into consideration. The model can well describe the robot's dynamic performances and the modelling approach used here can be extended to other dielectric elastomer actuators with complicated geometries for control purposes. The obtained model allows us to design a feedforward plus feedback control scheme for the robot to achieve desired motion. Simulation shows fast response and good tracking performances which are further confirmed by the experiments.

In this study, a new design approach for a multi terrain-abilities crawling robot is presented. With a slender body and four legs, the robot is able to perform both Serpentine and quadruped locomotion. To generate the locomotion, a central pattern generator (CPG) inspired controller is adopted. According to biology studies, CPGs are a set of neuronal circuits, which are responsible for producing rhythmic motion employed in animal locomotion. The main feature of proposed locomotion generation approach is the use of a set of coupled Hopf Oscillators to imitate of CPG in a nerve system. Moreover, to deal with dynamically changing environments, a feedback based fuzzy logic control strategy is investigated. Finally, the effectiveness of the proposed robot system is verified through the experiments of a crawling robot prototype.

In this paper, a comparative study of speed control efficiencies based on two dynamical models is presented. The common structure of dynamical models is developed using first principles of physics. The complex thrust mechanism is studied using two different techniques. First, we apply a novel data-driven approach to build up the non-linear mapping between tail angular motion and thrust generated, which is essential to specify the input gradient. Further, additional experiments are designed to acquire data necessary for modelling the input delay factor causing a time-delayed input to drive the system output. Second, analytically developed Lighthill's slender body theory is also employed to study thrust mechanism for comparative purposes. Though the theory has been used in robotic fish motion studies, it was originally formulated as an analytical tool for studying biological fish motion. Comparing the two theories help understand the predominant features in a robotic fish speed control. Lastly, because of the highly non-linear system dynamics, robust discrete-time Sliding Mode Controllers (SMC) are developed with the two dynamical models as basis respectively. Experimental verifications confirm that SMC based on data-driven model performs superior to the SMC based on the Lighthill's theory, by efficiently controlling the robotic fish's speed to track the time-varying reference speeds.

In this paper, a novel work is presented, where a learning-based control approach is proposed for motion control for a two-link robotic fish. First, by virtue of the Lagrangian mechanics method, we establish a mathematical model for the two-link Carangiform robotic fish. According to the constructed dynamical model, P-type learning control laws are proposed for speed and turning control of the robotic fish. Furthermore, due to the complexity of the dynamical model of the robotic fish, a self-adaption rule is introduced for learning gains, which might expedite the convergence rate of learning. In the end, the efficiency of the proposed learning controllers are illustrated by simulations.

This paper investigates motion control of a biomimetic robotic fish through a general internal model (GIM)-based learning approach. By virtue of the universal function approximation ability and the temporal/spatial seal-abilities of the GIM, the learning approach is able to generate the same or similar fish swimming motion patterns. Through experimental analysis, we also find out that motion of the robot can be controlled by simply tuning two GIM parameters. To achieve desired motion, based on tuning the GIM parameters, two model-free feedback control schemes are designed. A proportional-integral-derivative (PID) controller and a Fuzzy logic controller (FLC) are employed to control the speed and the turning diameter of the robot, respectively. Finally, experiment results verify the effectiveness of the motion control approach.

This paper presents a novel biomimetic learning approach for a Carangiform robotic fish to learn swimming locomotion. A video recording system is first set up to capture real fish behaviors that are used as the training samples. Three basic Carangiform swimming motion patterns, “cruise”, “cruise in turning” and “C sharp turn”, are extracted from robotic perspective. A general internal model (GIM) is adopted as a universal central pattern generator (CPG). Based on the universal function approximation ability and the temporal/spatial scalabilities of GIM, biomimetic learning is performed such that the robotic fish is able to learn to generate the same or similar fish swimming motion patterns. The three swimming motion patterns are implemented on a multi-joint robotic fish. The effectiveness of the biomimetic learning approach is verified through experiment results.

Rotor is the most important component in mini autonomous helicopter. In order to simplify identification and control, it is desirable for mini autonomous helicopter to hold rotor speed constant, since varying rotor speed causes varying aerodynamics. An intelligent rotor speed controller has been developed, which adopts the control algorithms of feed forward and fuzzy tuned PI. The hardware design especially the tachometer design is described in this paper. To improve the performance of PI controller, a fuzzy logic supervisor has been proposed for tuning the gain parameters of PI controller online. The controller is then applied in a mini autonomous helicopter to verify its performance. The experimental results of real flights have shown that the intelligent rotor speed controller is superior in dynamic and steady state performance for its advantages of quick response and good robustness. Numerous flights have also proven its reliability