This paper presents the design, modeling, control, and experimental validation of a biomimetic robotic fish that emulates thunniform locomotion. Motivated by the challenges of traditional underwater vehicles in terms of maneuverability and adaptability, the study aims to leverage bio-inspired propulsion strategies to enhance aquatic navigation in constrained environments. The fish robot is designed and developed with a hydrodynamic structure, incorporating active and passive fins and flexible joints to mimic natural fish movements. The dynamic and kinematic models are derived using Lagrangian mechanics, and a Central Pattern Generator (CPG)-based control scheme is implemented to generate rhythmic joint actuation without requiring precise trajectory planning. Simulation studies conducted in MATLAB/Simulink demonstrate smooth transitions from static to straight-line and turning maneuvers with a maximum lateral deviation of under 2 cm and a root mean square trajectory error of 0.0143 m. Experimental validations, both mid-air and in a water tank, confirm the sinusoidal motion patterns and verify the effectiveness of the control strategy. The results showcase the robot’s capability for stable and lifelike planar swimming, offering a promising platform for further developments in autonomous underwater systems.

This study introduces adaptive control for training pediatric cerebral palsy (CP) patients with an exoskeleton to improve gait. The patient-exoskeleton dynamic model, complex due to uncertainties, poses challenges for effective gait-tracking control. To address this, we propose radial basis function-based adaptive non-singular fast terminal sliding mode (RBF-ANFTSM) control for passive-assist exoskeleton systems. The proposed control mitigates uncertainties by leveraging radial basis function neural networks, while non-singular fast terminal sliding mode control ensures finite-time convergence and tackles singular issues. The Lyapunov approach, with adaptive weight laws, ensures system stability and robustness. Effectiveness is demonstrated with a 12-year-old CP patient coupled with the exoskeleton. Comparative analysis reveals that the proposed adaptive control outperforms existing ANFTSM control and RBF-computed torque control by more than 60% and 50%, respectively, in terms of gait tracking efficiency. Additionally, the proposed control exhibits converging sliding surfaces with lesser chattering (|s|hipkneeankle< 0.05) and offers localized convergence behavior over each gait cycle in estimating lumped uncertainties due to the influence of sliding surfaces in the adaptive weight law, underscoring its effectiveness.

This study proposes an adaptive non-singular fast terminal sliding mode (ANSFTSM) control for an exoskeleton system in passive-assist mode. The exoskeleton serves the purpose of motion assistance and gait restoration for the pediatric group aged 8–12 years. The complexity of the dynamic model, characterized by uncertain dynamics and external perturbations, poses a challenge in developing a gait tracking control for a pediatric exoskeleton. The experimental setup and real-time control architecture are introduced, followed by the dynamic formulation of the subject-exoskeleton system. Thereafter, a non-singular fast terminal sliding surface (NSFTSM) is considered to design the proposed equivalent control scheme. The adaptive laws are introduced in the reaching control scheme. Lyapunov's theory is employed to verify the rapid convergence of the tracking error in a finite time. The proposed ANSFTSM control is implemented for a pediatric subject (12 years) coupled with the exoskeleton in passive-assist gait tracking. The robust control provides efficient gait tracking (RMSEhip: 2.06°, RMSEknee: 3.81°, RMSEankle: 1.26°) with converging sliding surfaces (|s|hipkneeankle< 0.04) and negligible chattering in finite-time. This study showcases the potential to apply the proposed control to a broader range of pediatric subjects.

Deep Reinforcement Learning has enabled the learning of policies for complex tasks in partially observable environments, without explicitly learning the underlying model of the tasks. While such model-free methods do achieve considerable performance, they often ignore the structure of task. We present a more natural representation of the solutions to Reinforcement Learning (RL) problems, within 3 Recurrent Convolutional Neural Network (RCNN) architectures to better exploit this inherent structure. The forward passes of each RCNN execute an efficient Value Iteration, propagate beliefs of state in partially observable environments, and choose optimal actions respectively. Applying back-propagation to these RCNNs allows the system to explicitly learn the Transition Model and Reward Function associated with the underlying MDP, serving as an elegant alternative to classical model-based RL. We evaluate the proposed algorithms in simulation, considering a robot planning problem. We demonstrate the capability of our framework to reduce the cost of re-planning, learn accurate MDP models, and finally re-plan with learned models to achieve near-optimal policies.