Online Learning of Joint-Muscle Mapping Using Vision in Tendon-driven Musculoskeletal Humanoids (2404.05295v1)

Abstract: The body structures of tendon-driven musculoskeletal humanoids are complex, and accurate modeling is difficult, because they are made by imitating the body structures of human beings. For this reason, we have not been able to move them accurately like ordinary humanoids driven by actuators in each axis, and large internal muscle tension and slack of tendon wires have emerged by the model error between its geometric model and the actual robot. Therefore, we construct a joint-muscle mapping (JMM) using a neural network (NN), which expresses a nonlinear relationship between joint angles and muscle lengths, and aim to move tendon-driven musculoskeletal humanoids accurately by updating the JMM online from data of the actual robot. In this study, the JMM is updated online by using the vision of the robot so that it moves to the correct position (Vision Updater). Also, we execute another update to modify muscle antagonisms correctly (Antagonism Updater). By using these two updaters, the error between the target and actual joint angles decrease to about 40% in 5 minutes, and we show through a manipulation experiment that the tendon-driven musculoskeletal humanoid Kengoro becomes able to move as intended. This novel system can adapt to the state change and growth of robots, because it updates the JMM online successively.

• The paper introduces a neural network method for online updating of joint-muscle mapping to address errors in tendon-driven humanoids.
• It employs an antagonism updater that reduces muscle tension discrepancy by about 120 Newtons across trials.
• The vision updater cuts shoulder RMSE from 16° to 3° by integrating real-time visual feedback for improved movement accuracy.

Online Learning of Joint-Muscle Mapping in Tendon-driven Musculoskeletal Humanoids

The paper "Online Learning of Joint-Muscle Mapping Using Vision in Tendon-driven Musculoskeletal Humanoids" explores the complexities inherent in tendon-driven musculoskeletal humanoids. These humanoids are modeled after human anatomy which imposes challenges in achieving accuracy comparable to traditional humanoid robots actuated by motors at each joint. Conventional geometric modeling of such humanoids introduces significant errors, notably leading to undesirable muscle tension and slack in tendons.

To address these challenges, the authors propose a methodology for accurately learning and updating the joint-muscle mapping (JMM) using real-time data from the robotic system. This mapping, which represents the nonlinear relationship between joint angles and muscle lengths, is typically done through polynomial regression or table-searching methods. However, the paper introduces the utilization of neural networks (NN) for expressing JMM, capitalizing on the NN's ability to facilitate online updates—a key feature lacking in previous methods.

The approach includes initial training from the geometric model followed by continuous refinement using actual robot data. The primary objective is to ensure the humanoid can move according to expected commands in real-world scenarios without having precise geometric models.

Methodology and Experiments

The investigation comprises the integration of two online learning mechanisms:

1. Antagonism Updater: This updater utilizes estimated joint angles alongside actual muscle lengths for continuous learning, effectively adjusting muscle antagonism and mitigating excessive internal muscle tension and tendon slack. Results indicated a significant reduction in muscle tension discrepancy of about 120 Newtons over the course of 11 trials with the elbow joint.
2. Vision Updater: Employing vision through an RGB camera to estimate actual joint angles, this strategy enables the humanoid to correct its movement orientation to the intended posture by assimilating data from visual estimations of hand or other parts placements. The outcome was a decrease in RMSE—initially at roughly 16 degrees for shoulder movements—down to approximately 3 degrees post-learning.

The combination of these methodologies illustrated a robust improvement in achieving targeted arm positioning, verified through extensive trials that revealed reductions in estimation errors significantly by 40%.

Practical and Theoretical Implications

The applicability of this research is profound, enabling tendon-driven musculoskeletal humanoids to perform manipulation tasks that were previously unattainable due to model discrepancies. By successfully integrating sensor feedback and learning capabilities, the paper paves the way for tendon-driven systems to mimic human movement more effectively, assuming a potential role in fields requiring biomimetic accuracy.

Moving forward, future work as envisaged by the authors involves enhancing real-time sensor recognition beyond AR markers, thereby increasing the autonomy of the humanoids. Moreover, incorporating muscle tension dynamics into the joint-muscle mapping would offer a finer control over musculoskeletal stiffness, promising better handling and precision during interactive robot tasks.

The paper contributes an incremental yet significant advancement in the field of humanoid robotics, specifically focusing on those mimetic systems that echo the functionality of human musculoskeletal architecture. The described approach could mark a step towards making these systems competitive and perhaps surpass traditional rigid-motor actuated humanoids in diverse applications.