Dynamic3D · Constrained cupboard¶
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✗ 1h 00m | 10-04 06:57 | 1h 00m | — | 13.4M / 284k | $0.333 | deterministic 1, grader_error 0, missing_trajectory 0, n_actions 933, video_rendered 1 |
| L✗ 23m | 10-04 07:00 | 23m | — | 14.4M / 125k | $0.225 | deterministic 1, grader_error 0, live_success 0, missing_log 0, n_actions 1000, 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✓ 9m | 10-01 06:00 | 9m | 35 | 1.8M / 12k | $0.430 | $0.466 | deterministic 1, grader_error 0, missing_trajectory 0, n_actions 974, video_rendered 1 |
| L✓ 45m | 10-01 02:52 | 45m | 105 | 7.8M / 72k | $1.78 | $1.86 | deterministic 1, grader_error 0, live_success 1, missing_log 0, n_actions 988, replay_success 1, video_rendered 1 |
Task instruction (upstream)
Put the six brown rods lying on the floor (30 cm long) into the six 40 cm-tall compartments of the row of eleven narrow cupboards standing side by side, one rod each, as the state names pair them: rod cuboid_i into cupboard cupboard_i for i = 0 to 5, whose tall compartment is the upper one in cupboard_0 to cupboard_2 and the lower one in cupboard_3 to cupboard_5. The task succeeds the moment every rod's centre is inside its compartment: above that compartment's shelf, below the shelf over it, and within the cupboard's 10 cm width and 30 cm depth.
What KinDER states about this task
| Success criteria | 1. KinDER's own goal check (the environment's terminated) fires within the episode: at most 1000 control steps (100 s at 10 Hz); 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 six brown rods lying on the floor (30 cm long) into the row of eleven narrow cupboards standing side by side ahead of the robot, one rod each, each into a cupboard's 40 cm-tall compartment: taking the rods in order of their distance from where the centre of the robot's base starts, nearest first, they go into the 11th, 5th, 2nd, 8th, 4th and 1st cupboard of the row, counted from the robot's left as it faces the row at the start. The task succeeds the moment every rod's centre is inside its compartment: above that compartment's shelf, below the shelf over it, and within the cupboard's 10 cm width and 30 cm depth. |
| Env Id | kinder/ConstrainedCupboard3D-o6-v0 |
| Robot | TidyBot++: holonomic base, Kinova Gen3 7-DoF arm, Robotiq 2F-85 gripper |
| Category | Dynamic3D (MuJoCo) |
| 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 | base pos and yaw (3), arm joints (7), gripper pos (1); each base and joint delta at most 0.1 a step |
| State Dim | 195 |
| Limited Mode | base RGB and wrist RGB-D cameras (480×640), base odometry, the arm's joint angles and velocities, the fingers' position. 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 @ ae0d92a50.
Tags¶
Why this task is interesting¶
Six 30 cm rods go into six compartments 10 cm wide and 30 cm deep, each into the one it is owed. Gripped in the middle, the wrist reaches the opening before the rod is in, so each rod has to be held by one end and turned to slide in along the cupboard's depth, six times within 1000 steps.
Capability notes¶
Not yet written.
Oracle demo review¶
No demo.
Discussion¶
GPT-6.1 Sol passed both modes, limited mode at step 988 of its 17th episode after finding, in its ninth, that a tilted grasp and a base push get the rods into the bays. GPT-6 Luna (2026-10-04) had five of the six rods in place in unlimited mode when the hour ran out, and its single limited episode ran its 1000 steps without success. (@williamzhangNU)