The rating of perceived exertion (RPE) and surface electromyography (sEMG) describe exercise intensity subjectively and objectively, while there has been a lack of research on the relationship between them during dynamic contractions to predict exercise intensity, comprehensively. The purpose of this study was to establish a model of the relationship between sEMG and RPE during dynamic exercises. Therefore, 20 healthy male subjects were organized to perform an incremental load test on a cycle ergometer, and the subjects’ RPEs (Borg Scale 6–20) were collected every minute. Additionally, the sEMGs of the subjects’ eight lower limb muscles were collected. The sEMG features based on time domain, frequency domain and time–frequency domain methods were extracted, and the relationship model was established using Gaussian process regression (GPR). The results show that the sEMG and RPE of the selected lower limb muscles are significantly correlated (p < 0.05) but that they have different monotonic correlation degrees. The model that was established with all three domain features displayed optimal performance and when the RPE was 13, the prediction error was the smallest. The study is significant for lower limb muscle training strategy and quantification of training intensity from both subjective and objective aspects, and lays a foundation for sEMG further applications in rehabilitation medicine and sports training.

Robotic hand has vast application potential in home-serviced, medical care, surgery, maintenance, Etc. Various robotic hands made of rigid and flexible materials have been developed in the past decade. In this work, we present a rigid-flexible coupled underactuated hand with good agility and compliance for various kinds of grasping and safe operations. First, the dynamic simulation for finger bending and underactuated hand grasping was studied. Then, a prototype of the underactuated hand was fabricated. Next, we carried out a series of experiments based on the prototype. Single finger bending performance and blocking force were tested, respectively. We also tested the adaptive grasping ability of the hand. Finally, we performed collision experiments and compared the impact forces of the Be Bionic hand and our underactuated hand. Results showed that the underactuated hand could grasp objects with different shapes and possessed a higher degree of safety, reducing impact force by 30% to 50%.

This paper presents a reward design method for peg-in-hole insertion in uncertain environments using low-dimensional observation of robot end-effector forces. Recent reinforcement learning (RL) studies often avoid reward design by incorporating prior knowledge, which increases data requirements for RL training. Other approaches use sparse rewards to reduce human-engineering effort but fail to provide non-zero values frequently enough to avoid wandering traps. We propose a reward design method that offers improved performance by utilizing raw 3-dimensional force observations and action elements defined through Markov Decision Process (MDP), instead of the commonly used vision-based observations. The training framework employs a policy-based RL algorithm. We compare our proposed reward with various penalty and encouragement settings, trained with different scale parameters. Additionally, we compare our method with multi-modal observation approaches, demonstrating advantages in stability and reduction of wandering steps, achieving an average success rate of approximately 80% in uncertain position insertion tasks, trained using Proximal Policy Optimization (PPO).