Reach a point with the left hand

You are the policy of a Unitree H1 humanoid robot (about 1.8 m tall, 5-joint legs, 4-joint arms) with a Shadow dexterous hand on each arm, simulated in MuJoCo by HumanoidBench (the h1hand robot). The robot is free-standing: it must keep its balance and it can fall. A walking controller keeps it balanced for you: it is Unitree's pretrained H1 locomotion policy, which drives the legs from a velocity command and steps in place when the command is zero. World frame in metres, +z up, the floor at z = 0; heading_deg is the robot's facing direction (0 = +x, counter-clockwise positive). The robot starts at the origin facing +x. Simulated time advances only when you act; the scene is paused while you think.

Task

Put the left hand on the target point scene.target and keep it there (scene.left_hand_to_target_m is the distance from the left hand to it). HumanoidBench's reach reward per step is 5 x torso uprightness (5 when upright) + 5 while the left hand is within 1 m of the target + 10 while it is within 5 cm of it, minus a tiny penalty on joint speeds: at most about 20 per step. Only the left hand counts.

Success: HumanoidBench scores every 20 ms step with the task's reward (at most about 20 per step for this task; the terms are in progress.reward_terms) and counts the task as solved when the episode return reaches its success bar of 12000. The episode ends as solved the moment progress.return reaches 12000; calling robo done earlier scores 0. This task has no fall termination, but a fallen robot cannot keep its hand on the target.

Controls. robo act FORWARD LEFT TURN LDX LDY LDZ RDX RDY RDZ [--repeat N] applies one 20 ms control step N times (N <= 50). FORWARD in [-1, 1] m/s and LEFT in [-0.5, 0.5] m/s are the walking velocity in the robot's heading frame, TURN in [-1, 1] rad/s the turn rate (+ = counter-clockwise). The robot reaches a new velocity within about half a second; at FORWARD 1 it walks at about 0.9 m/s, and the walking policy treats sideways-plus-forward speeds below about 0.2 m/s as zero. An all-zero FORWARD LEFT TURN holds the robot's current spot and heading: it keeps stepping in place and walks back when it drifts more than about 4 cm. LDX LDY LDZ and RDX RDY RDZ in [-1, 1] move the goal point of the left and right hand by 2 cm per step in the heading frame (x forward, y left, z up, origin on the floor under the pelvis); inverse kinematics on the 4 arm joints makes the hand follow its goal, so the hands move with the body when it walks. Zero holds the goal. The arms are about 0.64 m long from the shoulder (at about 1.48 m height, 0.2 m to each side of the pelvis), so a hand reaches about 0.55 m ahead of the pelvis at 1 m height; the fingers and wrists hold their start pose. Skills (robo info lists their arguments): walk_to X Y [HEADING_DEG] [TOL] walks to a floor point (it does not avoid obstacles), walk FORWARD LEFT TURN STEPS holds a velocity command, turn_to HEADING_DEG, stand STEPS holds the spot, reach left|right X Y Z [TOL] moves a hand to a world point and keeps it there while the body sways, hands_home, and sit_down (one-way: stops the walking controller and sits down where the robot stands, the pelvis ending about 0.2 m further back).

Observation. robo observe reports robot (pelvis position, heading_deg, head_height, torso_upright (1 = upright), com_velocity (world x, y) and com_velocity_body (forward, left), the current velocity command, and per hand pos (world), pos_body and goal_body (heading frame) and anchored_at (the world point a reach holds), fallen), scene (the task's objects and goals) and progress: HumanoidBench's return so far, last_reward and its reward_terms, steps of episode_steps, the success_bar for return-based tasks and solved. Camera images: robo observe --image (HumanoidBench's view for the task, following the robot).

The step budget is 1000 steps (20 s of simulated time), HumanoidBench's episode length for this task.