In this page you will find some videos related with reinforcement learning with Dynamic Movement Primitives
Compliant admittance-coupled DMP without Type II singularity evader
The original reference trajectory is static in position and zero in force, so any external force triggers the movement of the mobile platform. However, in this video, the authors want to emphasize the problem of this kind of scheme in parallel robots: the control of the human upon the robot because of the compliant manipulation may lead the robot to a singular configuration and a subsequent control loss. Indeed, the original trajectory is close to a singularity in the axis of rotation around Z (ψ), so a moment on this axis (bottom-right chart) may cause its movement (bottom-left chart) to enter a singular configuration (which is measured by minΩ_c in the top chart), and the robot eventually responds unpredictably. The solution for this is the incorporation of a Type II singularity evader. In the charts, “ref” is the reference of the rotation around Z, “meas” is the measurement of the signals, and “lim” is the limit for minΩ_c.
The core controller is a PD+G, used to track the trajectory generated by the DMP.
Compliant admittance-coupled DMP with Type II singularity evader
The original reference trajectory is static and initially close to a singularity in the axis of rotation around Z (ψ). In this experiment, the human tries to push the robot toward the singularity by exerting a torque on Z (bottom-right chart), which provokes its rotation thanks to the admittance behavior (bottom-left chart). However, as soon as the minΩ_c singularity indicator falls below the limit, the evader is activated through a coupling action and keeps the robot out of the singularity, also ceasing the admittance behavior gradually. This results in a struggle between the human and the robot with smooth movements to ensure their safety. In the charts, “ref” is the reference of the rotation around Z, “meas” is the measurement of the signals, and “lim” is the limit for minΩ_c.
The core controller is a PD+G, used to track the trajectory generated by the DMP, affected by both coupling actions.
Type II singularity evader effect on a singular trajectory encoded with DMP
This video shows the execution of a trajectory generated by Dynamic Movement Primitives (DMP) to avoid Type II singular configurations for the 4-DOF parallel robot. To this end, an initial singular trajectory is encoded by the DMP, which is modified during execution by using coupling actions.
Those coupling actions are activated when either the determinant of the forward Jacobian ||J_D || or the minimum angle between two OTS minΩ_c fall below a threshold. Their values are obtained using a vision system to measure the cartesian pose, and only the two limbs most responsible for the singularity are affected by the coupling action. In this experiment, limbs 3 and 4 are deviated from their original reference (bottom charts) to keep the minΩ_c in the limit (top chart). In the charts, “ref” is the reference of the joint positions, “meas” is the measurement of the signals, and “lim” is the limit for minΩ_c.
A core PD+G controller tracks the modified trajectory free of singularities. The coupling actions are defined through objective controls on the variables ||J_D || and minΩ_c.
Imitation Learning-base System for the Execution of Selft-pace Robotic-assited Passive Rehabilitation Exercises
The development of robotic-assisted rehabilitation exercises involving physical human-robot interaction requires extreme care since an injured limb may be in physical contact with the robot, so compliant behavior is imperative for these tasks. Typical approaches involve force control schemes like admittance controllers that allow humans to adapt the motion. However, if motivated by a knee-jerk reaction, the force exerted by the human can be large and badly oriented, with the opposite effect to that intended.
This videos shows a new way of generating compliant trajectories for passive rehabilitation exercises, considering that previous positions of the trajectory are attainable for the patient, so reversing the trajectory is a safe operation. Since there is no clear way to optimize such a goal due to the physiological variability among patients, the condition of reversal is based on imitation learning by taking the analogous healthy limb of the patient as a reference and encoding the forces using Gaussian Mixture Regression, and reversibility is accomplished by means of Reversible Dynamic Movement Primitives. The system allows for self-paced rehabilitation exercises by back-and-forth movements along the trajectory according to the patient’s reaction, and it has been successfully applied to a 4-DOF parallel robot for knee rehabilitation
