JamesK2W's picture
LabUtopia masters: 9 task videos from GPT-6 Luna's counted privileged successes (owner decision; replays end on the grader's state hash) + teaser + highlights
bc87465 verified
Raw History Blame Contribute Delete
2.88 kB
{
"suite": "labutopia",
"task": "open_door",
"task_dir": "labutopia-open-door-i00-privileged",
"source": "agent-success:codex-gpt6_luna-xhigh-labutopia-open-door-i00-privileged/labutopia-open-door-i00-privileg__Vr32mjY",
"agent": "codex + GPT-6 Luna (xhigh)",
"run": "codex-gpt6_luna-xhigh",
"mode": "privileged",
"trial": "codex-gpt6_luna-xhigh-labutopia-open-door-i00-privileged/labutopia-open-door-i00-privileg__Vr32mjY",
"trial_choice": "only GPT-6 Luna solved this task (counted success); by the owner's decision (2026-10-23) its privileged trial is shown instead of the reference solution",
"renderer": "isaac-rtx PathTracing spp=32 (totalSpp=32) denoiser=optix maxBounces=6 fireflyFilter=on; rendered natively at 2560x1440 from a dedicated USD render camera (none of LabUtopia's observation cameras); Isaac Sim 6.1.0 / LabUtopia @ 8df7278",
"rendering": "path-tracing",
"media_standard": "v2 (2026-10-16)",
"width": 2560,
"height": 1440,
"fps": 30,
"frames": 722,
"seconds": 24.07,
"speed": 1.0,
"success_check": true,
"success_check_how": "the task's own rule (rcb_labutopia Session.status() success) at the end of this render's own Session.replay() of the trial's submitted trajectory (a fresh process: LabUtopia's scene and reset with the instance's seed, CPU PhysX, one row per 1/60 s control step until the rule ended the episode, frozen tests/instance.json), as tests/grade.py replays; the end-state hash equals the grader's recorded state_sha (verifier grade.log, both grading replays) and a separate plain replay with the tooling's own CLI (rcb_labutopia.py replay, fresh process, no renderer)",
"end_check": {
"success": true,
"rule": "first",
"reason": "success",
"score": 1.0,
"sim_time_s": 22.033333
},
"replay": {
"trajectory_sha256": "ac5cae6c550c9a4018553b4e74a78ed84eaabdd2a08230c4eba248621a86fd93",
"n_actions": 1323,
"control_steps": 1322,
"recorded_grader_steps": 1322,
"control_hz": 60,
"every_control_steps": 2,
"reason": "success",
"state_sha": "be87e4876c4c962f",
"recorded_grader_state_sha": "be87e4876c4c962f",
"matches_recorded_grader": true,
"cli_replay_state_sha": "be87e4876c4c962f",
"matches_cli_replay": true,
"instance_ok": true,
"start_exact": true,
"trajectory": "artifacts/app/output/trajectory.npz (the trial's submitted trajectory)"
},
"camera": {
"mode": "fixed",
"position": [
-1.1385,
-1.4418,
2.6329
],
"target": [
-0.0161,
0.1612,
1.104
],
"focal_mm": 35.0,
"h_aperture_mm": 36.0
},
"holds": {
"start_s": 0.5,
"end_s": 1.5,
"note": "first and last rendered states held (frames cloned), no extra simulation"
},
"poster": "last frame (solved state)",
"seconds_per_frame": 3.685,
"image": "efe-private:labutopia-0.1.0",
"encode": "libx264 preset slow CRF 16, yuv420p, bt709, faststart",
"created": "2026-10-23T10:02:39+0000"
}