Upper limb prostheses are commonly propelled by pneumatic artificial muscles organized in an antagonistic arrangement. Nonetheless, the control of upper limb prostheses under changing/unknown situations is difficult and necessary for a variety of real-world applications. Adaptive control, learning-based control, and robust control have been studied to deal with such challenges. However, their adaptability is insufficient for prostheses used in daily life, which are exposed to variable task levels, user motor characteristics, and prosthetic features. This paper introduces a highly adaptive controller for the first time based on Generative Adversarial Nets and proportional–integral–derivative controller (G-PID controller). G-PID controller comprises a generator for generating compensation actions to enhance PID responsiveness when controlling the unknown/changing system. Moreover, it incorporates a discriminator that receives responses from both a user-preselected reference system and the compensated changing/unknown system, and simultaneously determines the source of these responses. Through continuous updates, the compensator modifies the response of unknown/changing system to align with the reference system, thereby facilitating adaptive control. The G-PID controller’s effectiveness is evaluated through 1-degree of freedom (DoF) joint and 2-DoF shoulder prostheses in simulation experiments, and further validated in prototype experiments focusing on online learning for unknown and time-varying payload. The results demonstrate its ability to deal with diverse types of unknowns/changes, marking a significant advancement towards incorporating prostheses seamlessly into daily life.

Pneumatic artificial muscles (PAMs) have been widely applied to robotic systems, especially assistive devices, which could benefit from PAMs’ intrinsic viscoelasticity. However, nonlinearity and hysteresis make it challenging to achieve high-accuracy control. Using pre-tensioned springs to replace one side PAMs in an antagonistic structure is one efficient way to improve control accuracy. However, in the previous study, the selection method for spring, especially the range of spring constant has not been sufficiently investigated. This study designed a method to find the optimal spring constant based on gradient descent algorithm; and verified it with prototype and simulation experiments. Results showed that the group with optimal spring constant leads to a faster response time, smaller overshoot, and lower steady-state error, which indicated that the proposed method could find optimal spring constant. Moreover, the sensitivity of PAMs dynamic model parameters to the optimal spring stiffness was investigated. This study provides insights to practical use of soft actuators.

Soft actuators, made from soft materials, offer a safer alternative to rigid robots for use on hand rehabilitation devices. A current challenge is to ensure these actuators comply with human finger morphology. To gain better insights into actuator mechanics when worn on and interacting with human fingers, combining physical experiments with simulation approaches is necessary. However, no simulation has been implemented for finger-actuator interactions. This letter proposes a new joint modular soft actuator designed to comply with a dummy finger joint. The new actuator has a dual function design for increasing axial elongation during bending, facilitating compliance with finger morphology. In addition, a novel FEM for the new actuator's interaction with the dummy finger joint (dummy joint-soft actuator complex) is developed and used with physical experiments for evaluating actuator performance. Results show that the new design increases the dummy joint's bending range while exerting smaller contact forces on the joint. Even when the joint is blocked at specific bending angles, the actuator remains compliant to finger morphology. This research is a significant advancement in soft actuator design for hand rehabilitation, emphasizing the interaction between human fingers and soft actuators.

To provide a stable surgical view in Minimally Invasive Surgery (MIS), it is necessary for a flexible endoscope applied in MIS to have adjustable stiffness to resist different external loads from surrounding organs and tissues. Pneumatic soft actuators are expected to fulfill this role, since they could feed the endoscope with an internal access channel and adjust their stiffness via an antagonistic mechanism. For that purpose, it is essential to estimate the external load. In this study, we proposed a neural network (NN)-based active load-sensing scheme and stiffness adjustment for a soft actuator for MIS support with antagonistic chambers for three degrees of freedom (DoFs) of control. To deal with the influence of the nonlinearity of the soft actuating system and uncertainty of the interaction between the soft actuator and its environment, an environment exploration strategy was studied for improving the robustness of sensing. Moreover, a NN-based inverse dynamics model for controlling the stiffness of the soft actuator with different flexible endoscopes was proposed too. The results showed that the exploration strategy with different sequence lengths improved the estimation accuracy of external loads in different conditions. The proposed method for external load exploration and inverse dynamics model could be used for in-depth studies of stiffness control of soft actuators for MIS support.

Robotic tactile learning is a critical research area for enabling robots to perform complex manipulation tasks with human-like dexterity and adaptability. However, ensuring grasp stability remains one of the most fundamental yet challenging problems in tactile sensing. Existing approaches predominantly focus on classifying stability outcomes, lacking the ability to reason about failure modes or propose corrections. We address this gap by harnessing the inherent reasoning capabilities of Vision-Language Models (VLMs). To do so, we introduce a framework that integrates tactile modality via Quantized Low-Rank Adaptation (QLoRA), enabling parameter-efficient fine-tuning that empowers VLMs to both diagnose grasp instability and generate actionable, natural language corrections. To tackle tactile data scarcity, we develop an automated data collection and annotation pipeline leveraging state-of-the-art tactile sensor simulation, calibrated via a real-to-sim procedure to ensure robust transfer. Extensive experiments across 15 k simulated and 400 real-world trials demonstrate superior zero-shot sim-to-real generalization. In real-time physical evaluations, our approach achieves a 95% grasp success rate on unseen objects, significantly outperforming baselines.