Manipulation · Push¶
Excluded from the benchmark
not among the hardest (robot_coding_bench 2026-10-04): passed in both modes
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✓ 36m | 10-01 06:45 | 36m | — | 30.0M / 131k | $0.387 | deterministic 1, grader_error 0, n_actions 9, video_rendered 1 |
| L✓ 24m | 10-01 14:47 | 24m | — | 8.1M / 55k | $0.120 | deterministic 1, grader_error 0, live_success 1, missing_log 0, n_actions 100, replay_success 1, 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 02:40 | 1h 00m | 119 | 9.9M / 88k | $2.16 | $2.24 | deterministic 1, grader_error 0, n_actions 8, video_rendered 1 |
| L✓ 14m | 10-01 00:02 | 14m | 43 | 2.2M / 24k | $0.625 | $0.670 | deterministic 1, grader_error 0, live_success 1, missing_log 0, n_actions 301, replay_success 1, 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✓ 4m | 09-28 19:48 | 4m | — | 863k / 14k | $0.992 | deterministic 1, grader_error 0, n_actions 142, video_rendered 1 |
| L✓ 13m | 09-28 17:34 | 13m | — | 5.1M / 43k | $4.09 | deterministic 1, grader_error 0, live_success 1, n_actions 403, replay_success 1, 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✓ 18m | 09-28 17:51 | 18m | — | 4.8M / 26k | $1.41 | deterministic 1, grader_error 0, live_success 1, n_actions 110, replay_success 1, video_rendered 1 |
Task instruction (upstream)
Push the box on the table to the target position marked in the observation.
What HumanoidBench states about this task
| Success criteria | 1. the summed per-step reward over one episode reaches 700, 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 | h1-push-v0 |
| Robot | Unitree H1 (19 actuators) |
| Category | Manipulation |
| Capability Class | C6 · standing bimanual manipulation |
| Role | scored |
| Scoring | Every step here is negative: it costs the box's distance to the target, plus a smaller charge for the left hand's distance to the box. The only positive term is a one-off bonus, paid when the box lands within a few centimetres of the target — and that also ends the episode. So the bar is not reached by scoring well for a long time; it is reached by finishing, and finishing early, because every extra step subtracts. |
| Zero Action Return | -253.71 |
| Action Dim | 19 |
| 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 where the hand or the box is, not the target's coordinates, 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¶
The box is out of reach from where the robot stands: it has to bring its body closer, then push the box into a target a few centimetres wide as early as it can, since the reward counts the steps.
Capability notes¶
Not yet written.
Oracle demo review¶
No demo.
Discussion¶
The easiest scored task: passes both modes, limited mode in 15 episodes with an ankle lean and arm kinematics from spec(). (@williamzhangNU)