Index fingertip to a random point

You control MyoFinger, MyoSuite's musculoskeletal model of a single index finger (MuJoCo physics), only through its 5 muscles: extn (extensor), adabR and adabL (sideways, abduction/adduction), mflx and dflx (flexors). The finger's base is fixed; it has 4 joints: IFadb (sideways, -0.44 to 0.44 rad), IFmcp (knuckle flexion, -0.44 to 1.05 rad), IFpip and IFdip (middle and end joint flexion, 0 to 1.05 rad). Angles are in radians. The finger starts straight (all joints 0), pointing along +x; flexing bends it towards -z. World frame in metres, +z up. A muscle can only pull, and the flexors cross several joints.

Task

The target was drawn at random from MyoSuite's target box (fixed here by the episode seed). Move the fingertip to the target position (MyoSuite env myoFingerReachRandom-v0): IFtip (index tip) to [0.1915, -0.0313, 0.1738] (world frame, m). The targets are also drawn as coloured spheres. At the start the error norm is 0.152 m.

Success: MyoSuite's own solved check for this env: the Euclidean norm of all fingertip position errors stacked together (reach_error_norm, 3 numbers) must be below 0.0125 m (12.5 mm per fingertip, threshold), judged by the episode server from the simulated state after you call robo done and the 10-step settle. Simulated time advances only when you act (one step = 20 ms).

Controls. robo act E1 E2 ... [--repeat N] sets the excitation (neural drive) of every muscle, a number in [0, 1], in the order robo info lists them (the action names), and holds it for N steps of 20 ms. A muscle's activation follows its excitation with MuJoCo's first-order activation dynamics (about 10 ms rising, 40 ms falling); its force also depends on its length and speed. The excitations stay as you last sent them until you act again, and they are also what the muscles hold during the 10-step settle after robo done. Skills (robo info lists them): robo skill set_joint_targets JOINT=RAD [JOINT=RAD ...] [steps=60] [ramp=20] runs a muscle-space controller: every step it computes the joint torques of a PD law towards the targets (MuJoCo inverse dynamics), solves a bounded least-squares problem for the muscle activations in [0, 1] that best produce those torques, and sends the matching excitations as that step's action (the steps count against the budget like robo act). Joints you do not name keep their previous targets (at the start: the start pose). The targets move linearly from the current pose over RAMP steps; the skill stops after STEPS steps (at most 300) or once the joints are still, and reports the remaining joint-target error and the task's error. robo skill hold [steps=20] sets every target to the current angle and keeps the joints there with the same controller. The controller cannot do the impossible: joints coupled by shared muscles or blocked by contact can stop short of their targets, so check the result and adjust. robo skill reach_tips TIP=X,Y,Z [TIP=X,Y,Z ...] [steps=60] [ramp=20] (positions in metres, no spaces inside a position) first solves inverse kinematics for the named fingertips within the joint ranges, starting from the current pose, then runs set_joint_targets with the result; it also reports the kinematic residual per fingertip.

Observation. robo observe reports fingertips and fingertip_targets (positions, m), reach_error_norm and threshold; joint_angles_rad, joint_velocities_rad_s and joint_targets_rad (the controller's current targets) per joint; muscle_activations (one per muscle, in action order); contacts (pairs of bones or objects touching); time_s; and obs_vector, MyoSuite's own observation vector for this env (its observation keys, in order: qpos, qvel, tip_pos, reach_err, act; not printed by plain robo observe, shown with robo observe --json). robo observe --image saves a picture from a fixed camera (muscle paths are not drawn).

The step budget is 300 steps (6 s of simulated time).