Manipulation · Window¶
Excluded from the benchmark
not among the hardest (robot_coding_bench 2026-10-04): the best run reached 95% of the bar
Greyed and last in the task list, left out of the benchmark's numbers; its demo and agent runs stay (excluded: in state/tasks/humanoidbench.yml).
Agent runs
Each mode of this task runs once per run. Est. cost is tokens at list price (data/prices.yml), never a bill; see the HumanoidBench runs for every task of a run.
Codex CLI 0.157–0.159 + GPT-6 Luna, reasoning medium (OpenRouter)default run
| Trial | Started | Agent time | Model requests | Tokens in / out | Est. cost | Grade |
|---|---|---|---|---|---|---|
| U✗ 58m | 09-28 18:24 | 58m | — | 32.6M / 228k | $0.494 | deterministic 1, grader_error 0, n_actions 849, video_rendered 1 |
| L✗ 1h 00m | 10-01 15:28 | 1h 00m | — | 32.9M / 133k | $0.414 | deterministic 1, grader_error 0, live_success 0, missing_log 0, n_actions 100, replay_success 0, video_rendered 1 |
3 other run(s) of HumanoidBench
Codex CLI 0.159.2 + GPT-6.1 Sol, reasoning medium (OpenRouter)
| Trial | Started | Agent time | Model requests | Tokens in / out | Est. cost | Billed | Grade |
|---|---|---|---|---|---|---|---|
| U✗ 59m | 10-01 04:48 | 59m | 121 | 14.1M / 112k | $2.95 | $3.06 | deterministic 1, grader_error 0, n_actions 1000, video_rendered 1 |
| L✗ 51m | 10-01 00:28 | 51m | 123 | 12.0M / 78k | $2.32 | $2.41 | deterministic 1, grader_error 0, live_success 0, missing_log 0, n_actions 488, replay_success 0, video_rendered 1 |
Claude Code 2.1.283 + Claude Opus 5.5, reasoning medium (OpenRouter)
Not in this run: not in this run (the two dearer models ran five tasks, chosen to span locomotion and manipulation, easy to hard)
Codex CLI 0.157.0 + GPT-6 Sol, reasoning medium (OpenRouter)closed
Not in this run: not in this run (the two dearer models ran five tasks, chosen to span locomotion and manipulation, easy to hard)
Task instruction (upstream)
Grasp the wiping tool — it starts in the air just in front of the hands and falls within the first second unless it is held — and wipe the window with it, moving steadily up and down.
What HumanoidBench states about this task
| Success criteria | 1. the summed per-step reward over one episode reaches 650, HumanoidBench's own success bar (a total of rewards, not a number of steps; an episode is at most 1000 control steps) 2. unlimited: both fresh-process replays of the handed-in trajectory reach it and end in the same state 3. limited: the run passes the moment a live episode reaches it (the recorded episode replays to the same state); otherwise the last episode is graded |
| Env Id | h1hand-window-v0 |
| Robot | Unitree H1 with two Shadow hands (61 actuators) |
| Category | Manipulation |
| Capability Class | C6 · standing bimanual manipulation |
| Role | scored |
| Scoring | A large share of each step is paid only while the tool is in contact with the glass — while it is not, that share is zero, so the bar cannot be reached without contact. The rest pays for moving the tool vertically at a target speed on the order of half a metre per second, for keeping both hands near the tool, and for standing with the head about 40 cm from where it started. |
| Ends Early | The episode ends immediately if the pelvis drops or the tool falls, i.e. if it is dropped. |
| Zero Action Return | 1.56 |
| Action Dim | 61 |
| Control Rate Hz | 50 |
| Limited Mode | head cameras (RGB 256×256), joint angles and velocities; a pelvis IMU and a camera fixed in the room when the run turns them on. Not where the robot is in the room, not the tool's pose, not the reward |
| Agent Budget | 3600 s of wall clock per mode |
From https://github.com/carlosferrazza/humanoid-bench @ cb11890, as defined in our task definitions @ 7d6a7a44.
Tags¶
Why this task is interesting¶
Hold a wiper against the glass and move it up and down at about half a metre per second while standing, without letting the tool slip out of the hand.
Capability notes¶
Not yet written.
Oracle demo review¶
No demo.
Discussion¶
Close in unlimited mode, 32 short of the bar: the wiper stays on the glass for 866 steps but moves at about half the asked speed. (@williamzhangNU)