Fold both temples of a pair of glasses and put them into a box

A Franka Emika Panda arm stands at a table (table top about 0.93-1.0 m above the floor, raised by a random amount per scene) with a four-fingered Allegro robot hand (16 joints) mounted on its flange instead of a gripper. The hand is about 25 cm long and 10 cm wide; its fingers are index, middle, ring and thumb. World frame in metres, z up.

Goal: A pair of glasses lies open on the table next to an open box. Fold both temples (the arms of the glasses) shut and put the folded glasses into the box.

Success: Both temple hinge angles (left_temple_angle, right_temple_angle) are above 1.1 rad and the glasses' centre is inside the box's inner cavity (within 14.5 cm of the box centre along x and y, between 4.5 cm below and 2.75 cm above it), continuously for 50 steps (1 s). The episode ends as solved the moment this condition is met; robo done before that scores 0.

Task fields in robo observe: glasses_* (frame pose), left_temple_angle, right_temple_angle (rad; 0 = open) and open_box_*.

Controls. robo act takes 22 numbers (DX DY DZ RX RY RZ F0 .. F15). Per arm: DX DY DZ move the commanded hand pose by 2 cm per unit along world x, y, z (up to +-8 units = 16 cm per step); RX RY RZ rotate the commanded hand orientation by 0.2 rad per unit about the world x, y, z axes (up to +-3); F0 .. F15 change the commanded finger joint angles by 0.3 rad per unit (up to +-6): F0-F3 index, F4-F7 middle, F8-F11 ring, F12-F15 thumb. For index, middle and ring, joint 0 spreads the finger sideways and joints 1-3 curl it (positive = flex toward the palm); for the thumb, joint 0 swings it across the palm (opposition) and joints 1-3 rotate and curl it. Commands are targets: the arm's operational-space controller and the finger position servos track them, so the hand lags a fast command and stops where contact blocks it. robo info lists each joint's range (the commanded angle is clipped to it). One step is 20 ms of simulated time. Actions are deltas, so --repeat N applies the same change N times. robo move-to X Y Z moves the hand (the point hand_pos) toward a point without changing its orientation or fingers, and robo grip G sets the hand to a generic power grasp (G > 0 closes to fraction G of it, G < 0 opens all fingers, 0 keeps them); the grasp preset does not suit every object, so finger-level robo act control is often needed. The DX DY DZ GRIP form in the general instructions below does not apply: every robo act needs all 22 numbers.

Observation. robo observe reports for the arm: hand_pos and hand_quat_wxyz (measured flange position and orientation, quaternion w, x, y, z), hand_target_pos and hand_target_quat_wxyz (the commanded pose), finger_joints and finger_joint_targets (16 measured and commanded joint angles in radians, in F0..F15 order); for objects, <name>_pos (body origin), <name>_quat_wxyz, <name>_tilt_deg (angle of the body's z axis from vertical) and <name>_yaw_deg; plus the task fields listed above. robo observe --image saves a picture from the front camera; --camera picks another (front, top0 (overhead), left, right, handcam_rgb (wrist)). The hand's orientation matters: at the start the palm faces down; look at a picture before you reach.

The step budget is 2500 steps (50 s of simulated time). A human teleoperator needed 422 steps for this scene.