Figure 2. Synthetic proprioceptive characters dataset generation.

(A) Multiple examples of pen-tip trajectories for five of the 20 letters are shown. (B) The same trajectories as in A, plotted as time courses of Cartesian coordinates. (C) Creating (hand) end-effector trajectories from pen-tip trajectories. (Left) An example trajectory of character a resized to fit in a 10 × 10 cm² grid, linearly interpolated from the true trajectory while maintaining the true velocity profile. (Right) This trajectory is further transformed by scaling, rotating, and varying its speed. (D) Candidate starting points to write the character in space. (Left) A 2-link, 4 degrees of freedom (DoFs) model human arm is used to randomly select several candidate starting points in the workspace of the arm (right), such that written characters are all strictly reachable by the arm. (E) (Left to right and down) Given a sample trajectory in (C) and a starting point in the arm’s workspace, the trajectory is then drawn on either a vertical or horizontal plane that passes through the starting point. We then apply inverse kinematics to solve for the joint angles required to produce the traced trajectory. (F) (Left to right) The joint angles obtained in (E) are used to drive a musculoskeletal model of the human arm in OpenSim, to obtain equilibrium muscle fiber-length trajectories of 25 relevant upper arm muscles. These muscle fiber lengths and their instantaneous velocities together form the proprioceptive inputs.

Figure 2—video 1. Supplementary video.

Download video file ^{(434.8KB, mp4)}

Video depicts the OpenSim model being passively moved to match the human-drawn character ‘a’ for three different variants; drawn vertically (left, right) and horizontally (middle).