A scalable joint-space controller for musculoskeletal robots with spherical joints

Abstract

Abstract-In the long history of robotics research, the most prominent problem has always been, to develop robots that can safely operate in human-centered environments. One way towards the goal of a safe, and human-friendly robot, is to incorporate more and more of the flexibility that can be found in humans, by mimicking the internal mechanisms. In this work we propose a scalable joint-space control scheme based on computed torque control for an anthropomimetic robot. To achieve this, the dynamic system model of the robot is decomposed into hierarchical subsystems, using scalable modeling algorithms where possible. Machine learning techniques were employed to tackle the problem of muscle force to joint torque mapping.The developed control scheme has been evaluated using the highly refined simulation of an anthropomimetic robot arm featuring 11 muscles, a revolute elbow joint and a spherical shoulder joint. We show trajectory tracking based on a lowlevel muscle and a high-level joint control scheme, taking into account the coupling between the joints due to inertial reactions and bi-articular muscles.

Computed muscle control for an anthropomimetic elbow joint

“…Robotic controllers on the other hand lack the scalability for more complex robots. The controller proposed in [12] has proven useful to control the simulation of an anthropomimetic robot arm and will in this work be evaluated for controlling a physical robot based on the control architecture described in [13].…”

Adaptive neural network dynamic surface control for musculoskeletal robots

“…The mapping of the derivative of a quaternion and the corresponding rotational velocities is well known [11]. We define a matrix A(α), as a diagonal matrix, in which each revolute joint is represented by a 1 and each spherical joint is represented by the corresponding mapping, such that [12]:…”

“…The system states of (11) and (12) have been replaced by x 1 and x 2 which denote the muscle force f and the actuator velocityθ, respectively. Note that the system state of the first equation…”

“…Within the backstepping technique, this requires the calculation of the time derivative of φ 1 and the consideration of interaction terms in the combined dynamics of x 0 and x 1 through a stepped Lyapunov analysis. In the next step, the dynamics of x 2 are controlled using equation (12). This again requires time derivatives in particular for φ 2 .…”

Toward Anthropomimetic Robotics: Development, Simulation, and Control of a Musculoskeletal Torso

“…The principle of virtual work can then be used to derive the missing transformation between the muscle forces f and joint torques H [17]: However, the analytical derivation of the muscle Jacobian for our anthropomimetic robots is not possible, because of the absence of CAD data, and because of the presence of tendon-skeleton collisions. Therefore, we decided to approximate it by (i) utilizing machine learning techniques to learn the geometric mapping between tendon lengths l and joint angles q, and (ii) differentiating this geometric mapping using the difference quotient [29]. As samples can be drawn from the entire joint space, the function approximator need only support the interpolation between samples, without the need for extrapolation.…”