Manipulation · Cube¶
Excluded from the benchmark
not among the hardest (robot_coding_bench 2026-10-04): passed in unlimited mode
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 17:17 | 58m | — | 33.2M / 137k | $0.446 | deterministic 1, grader_error 0, n_actions 500, video_rendered 1 |
| L✗ 36m | 10-01 14:17 | 36m | — | 15.8M / 126k | $0.238 | deterministic 1, grader_error 0, live_success 0, missing_log 0, n_actions 89, 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✓ 1h 00m | 10-01 02:41 | 1h 00m | 121 | 10.6M / 99k | $2.37 | $2.45 | deterministic 1, grader_error 0, n_actions 500, video_rendered 1 |
| L✗ 1h 00m | 09-30 22:37 | 1h 00m | 171 | 19.5M / 100k | $3.37 | $3.48 | deterministic 1, grader_error 0, live_success 0, missing_log 0, n_actions 491, replay_success 0, video_rendered 1 |
Claude Code 2.1.283 + Claude Opus 5.5, reasoning medium (OpenRouter)
| Trial | Started | Agent time | Model requests | Tokens in / out | Est. cost | Grade |
|---|---|---|---|---|---|---|
| U✗ 17m ⓘ | 09-28 19:50 | 17m | — | 1.5M / 23k | $1.57 | deterministic 1, grader_error 0, n_actions 500, video_rendered 1 |
| L✗ 6m ⓘ | 09-28 17:35 | 6m | — | 1.4M / 20k | $1.44 | deterministic 1, grader_error 0, live_success 0, n_actions 182, replay_success 0, video_rendered 1 |
Codex CLI 0.157.0 + GPT-6 Sol, reasoning medium (OpenRouter)closed
| Trial | Started | Agent time | Model requests | Tokens in / out | Est. cost | Grade |
|---|---|---|---|---|---|---|
| Unot run | — | — | — | — / — | — | — |
| L✗ 11m | 09-28 17:55 | 11m | — | 1.7M / 22k | $0.668 | deterministic 1, grader_error 0, live_success 0, n_actions 198, replay_success 0, video_rendered 1 |
Task instruction (upstream)
A cube starts just above and in front of each palm. Rotate both cubes to the goal orientation without dropping them.
What HumanoidBench states about this task
| Success criteria | 1. the summed per-step reward over one episode reaches 370, HumanoidBench's own success bar (a total of rewards, not a number of steps; an episode is at most 500 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-cube-v0 |
| Robot | Unitree H1 with two Shadow hands (61 actuators) |
| Category | Manipulation |
| Capability Class | C7 · in-hand precision |
| Role | scored |
| Scoring | A weighted sum: the largest term is how closely each cube's orientation matches the goal, averaged over the two hands; then how close each hand is to its cube; then a small share for standing steadily without moving. The orientation term is sharp — it falls away quickly as the orientation drifts, so approximate alignment earns much less than it looks like it should — and it compares orientations as quaternions, sign included: of the two ways to turn a cube onto the goal, only one scores, because the other ends at the same orientation with the opposite sign, which counts as far off. |
| Ends Early | The episode ends the moment either cube drops below a height, or the pelvis does. |
| Zero Action Return | 5.0 |
| Action Dim | 61 |
| Control Rate Hz | 50 |
| Limited Mode | head cameras (RGB 512×512), 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 cubes' orientations, not the goal orientation, 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¶
Two Shadow hands must each turn a cube to a signed target orientation and hold it there. The starting pose does not stand by itself (it falls after about 84 steps), so the robot must first stay up for the 500-step episode, then turn the cubes early.
Capability notes¶
Not yet written.
Oracle demo review¶
No demo.
Discussion¶
Passes in unlimited mode with centre-of-mass balance and Jacobian wrist control turning both cubes; limited mode stays up and comes close (338.91 against 370) but drops a cube at step 491. (@williamzhangNU)