This paper proposes a systemic approach to upper arm gym-workout classification according to spatio-temporal features depicted by biopotential as well as joint kinematics. The key idea of the proposed approach is to impute a biopotential-kinematic relationship by merging the joint kinematic data into a multichannel electromyography signal and visualizing the merged biopotential-kinematic data as an image. Under this approach, the biopotential-kinematic relationship can be imputed by counting on the functionality of a convolutional neural network: an automatic feature extractor followed by a classifier. First, while a professional trainer is demonstrating upper arm gym-workouts, electromyography and joint kinematic data are measured by an armband-type surface electromyography (sEMG) sensor and a RGB-d camera, respectively. Next, the measured data are augmented by adopting the amplitude adjusted Fourier Transform. Then, the augmented electromyography and joint kinematic data are visualized as one image by merging and calculating pixel components in three different ways. Lastly, for each visualized image type, upper arm gym-workout classification is performed via the convolutional neural network. To analyze classification accuracy, two-way rANOVA is performed with two factors: the level of data augmentation and visualized image type. The classification result substantiates that a biopotential-kinematic relationship can be successfully imputed by merging joint kinematic data in-between biceps- and triceps-electromyography channels and visualizing as a time-series heatmap image.
Geduldspiele cubes (also known as patience cubes in English) are interesting problems to solve with robotic systems on the basis of machine learning approaches. Generally, highly dexterous hand and finger movement is required to solve them. In this paper, we propose a reinforcement-learning-based approach to solve simple geduldspiele cubes of a flat plane, a convex plane, and a concave plane. The key idea of the proposed approach is that we adopt a sim-to-real framework in which a robotic agent is virtually trained in simulation environment based on reinforcement learning, then the virtually trained robotic agent is deployed into a physical robotic system and evaluated for tasks in the real world. We developed a test bed which consists of a dual-arm robot with a patience cube in a gripper and the virtual avatar system to be trained in the simulation world. The experimental results showed that the virtually trained robotic agent was able to solve simple patience cubes in the real world as well. Based on the results, we could expect to solve the more complex patience cubes by augmenting the proposed approach with versatile reinforcement learning algorithms.
In consecutive upper-limb motor coordination tasks, where a new condition is imposed once a subgoal is achieved, kinematic demonstrations reveal what a human does but not how cautiously, so existing imitation learning misses the adaptation to the new condition. To close this gap, we propose designing reward and constraint functions from biopotential-informed characteristics of human surface electromyography (sEMG). The task begins as a Markov decision process (MDP) and turns into a constrained MDP (cMDP) once the first subtask is completed. The sEMG data are decomposed into shared and task-specific muscle synergies; the former infers a quadratic-regulator-formed reward by inverse optimal control, and the latter defines a subtlety coefficient bounding the action magnitude in the constrained phase. A two-holes two-balls patience cube exemplifies the task, and the policy is trained by constrained policy optimization and deployed on a robot manipulator. Whereas unconstrained PPO and SAC agents complete the constrained phase in only 1% of the episodes, the proposed agent completes 100% with the shortest completion time (1.76 s) among four constrained variants and most closely reproduces the human’s dwell-time distribution and zone transitions near the hole, yielding a feasible and human-like policy.