Kinematic3D · Obstruction 3D¶
Excluded from the benchmark
not among the hardest (robot_coding_bench 2026-10-04): GPT-6 Luna passed it in unlimited mode in 11 min
Greyed and last in the task list, left out of the benchmark's numbers; its demo and agent runs stay (excluded: in state/tasks/kinder.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 KinDER 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✓ 11m | 09-29 15:14 | 11m | — | 3.3M / 41k | $0.061 | deterministic 1, grader_error 0, missing_trajectory 0, n_actions 1306, video_rendered 1 |
| L✗ 1h 00m | 10-01 15:22 | 1h 00m | — | 34.0M / 153k | $0.462 | deterministic 1, grader_error 0, live_success 0, missing_log 0, n_actions 86, replay_success 0, video_rendered 1 |
1 other run(s) of KinDER
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✓ 3m | 10-01 05:55 | 3m | 16 | 546k / 4k | $0.192 | $0.216 | deterministic 1, grader_error 0, missing_trajectory 0, n_actions 1011, video_rendered 1 |
| L✓ 5m | 10-01 02:43 | 5m | 25 | 721k / 9k | $0.248 | $0.272 | deterministic 1, grader_error 0, live_success 1, missing_log 0, n_actions 428, replay_success 1, video_rendered 1 |
Task instruction (upstream)
Put the purple block (target_block) on the thin purple pad on the wooden table (target_region); three of the four small red blocks (obstruction0, obstruction1, obstruction3) sit on the pad and the fourth (obstruction2) is elsewhere on the table. The task succeeds the moment the robot holds nothing, the block's underside is within 5 mm of the pad's top surface, and all four of its bottom corners lie within the pad's outline enlarged by 5 mm on every side.
What KinDER states about this task
| Success criteria | 1. KinDER's own goal check (the environment's terminated) fires within the episode: at most 3000 control steps; no partial credit2. unlimited: both fresh-process replays of the handed-in trajectory reach it and end in the same state 3. limited: the one episode reaches it (no reset in our runs); the service records the episode, and it replays to the same state |
| Limited Instruction | Put the purple block on the thin purple pad on the wooden table; three of the four small red blocks sit on the pad and the fourth is elsewhere on the table. The task succeeds the moment the robot holds nothing, the block's underside is within 5 mm of the pad's top surface, and all four of its bottom corners lie within the pad's outline enlarged by 5 mm on every side. |
| Env Id | kinder/Obstruction3D-o4-v0 |
| Robot | TidyBot++, kinematic: holonomic base, Kinova Gen3 7-DoF arm, gripper; a move that would collide is not carried out |
| Category | Kinematic3D (PyBullet) |
| Instance | seed 0: the scene KinDER builds from it, pinned by its digest |
| Holding Still | does not reach the goal (checked by running an episode of zero actions) |
| Action Dim | 11 |
| Actions | bounded relative base position, rotation, and joint positions, and open / close; each base and joint delta at most 0.2 a step |
| State Dim | 91 |
| Limited Mode | base RGB and wrist RGB-D cameras (480×640), base odometry, the arm's joint angles, the fingers' position; a move that would collide comes back blocked. Not where any object or the goal region is, not the goal check before the episode ends |
| Agent Budget | 3600 s of wall clock per mode |
From https://github.com/Princeton-Robot-Planning-and-Learning/kindergarden @ 5b2dbac, as defined in our task definitions @ 7d6a7a44.
Tags¶
Why this task is interesting¶
Three small blocks sit on a pad barely larger than the block that must go there (7.8 × 5.6 cm for a 6.3 × 4.5 cm block), so the robot has to clear them first. The rules are kinematic: a move that would collide is not carried out, and a grasp fails if the fingers touch the pad or a neighbouring block.
Capability notes¶
Not yet written.
Oracle demo review¶
No demo.
Discussion¶
Both modes pass within minutes and without a reset: the limited run cleared the three red blocks with quarter-turned grasps placed by wrist depth. (@williamzhangNU)