masters: dunk_realgrasp = GPT-6 Astra's privileged success (BNuT7HK) on the 2.60 m ring, replaces the failed-attempt orbit

masters/roboolympics-fun/dunk_realgrasp/meta.json

@@ -1,77 +1,91 @@
 {
  "suite": "roboolympics-fun",
  "task": "dunk_realgrasp",
- "source": "scene-orbit",
- "why_no_video": "No evaluated agent has solved this task (no counted success in any run when this was made). As for the other unsolved RoboOlympics-Fun tasks on the site (board 0161 rule), this video shows the best failed attempt instead, replayed exactly (not a solved run): the task's own check at the end says it failed, as the verifier did.",
- "shows": "agent-attempt (failed):codex-gpt6_astra-high-roboolympics-fun-dunk-realgrasp-i00-privileged/roboolympics-dunk-realgrasp-i00__ZWR3UwT",
+ "source": "agent-success:codex-gpt6_astra-high-embodied-olympics-fun-dunk-realgrasp-i00-privileged/roboolympics-fun-dunk-realgrasp__BNuT7HK",
  "renderer": "Blender 4.2.23 LTS Cycles path tracing (GPU OptiX), adaptive sampling (threshold 0.01, 64-128 spp), OpenImageDenoise with albedo + normal passes; kinematic playback of the exact MuJoCo replay poses (MuJoCo's own compiled meshes, primitives and colours); daylight: a sun lamp (2 degree disc) and a Nishita sky (no sun disc), the venue's own ground colour on a large ground plane with a faint world-space variation; AgX view transform, Medium High Contrast look; a tracking camera (fixed 22 degree elevation, per-frame fit of the robot and the object in play, smoothed in time) with a slow eased 10-degree orbit",
  "rendering": "path-tracing",
  "width": 2560,
  "height": 1440,
  "fps": 30,
- "frames": 353,
- "seconds": 11.77,
- "speed": 0.3,
- "success_check": false,
- "success_check_how": "the task's own verdict (rcb_rfo.py Episode.verdict(), the referee, from the task dir's tests/ (privileged verifier copy)) at the end of this re-render's own replay of every recorded row from the task's frozen instance (t = 0 state hash checked); a second plain replay with the task's own replay() agrees and ends in the same final state hash (ad8cf0efa21c1183)",
+ "frames": 285,
+ "seconds": 9.5,
+ "speed": 0.6,
+ "success_check": true,
+ "success_check_how": "the task's own verdict (rcb_rfo.py Episode.verdict(), the referee, from the task dir's tests/ (privileged verifier copy)) at the end of this re-render's own replay of every recorded row from the task's frozen instance (t = 0 state hash checked); a second plain replay with the task's own replay() agrees and ends in the same final state hash (e6b841248f77092d), which is also the final state hash in the trial's own replay records (video_replay.json). Checked with both the task as deployed when the trial ran (the -lower variant, exam commit 671a3ebd42) and the exam's current main (3ac12a3e86, PR #448): same verdict, same final hash, byte-identical scene export",
  "end_check": {
-  "success": false,
-  "stages": [],
+  "success": true,
+  "stages": [
+   "dunk",
+   "stand"
+  ],
   "stages_total": 2,
-  "progress": 0.0,
+  "progress": 1.0,
   "metrics": {
-   "ball_height_m": 3.073,
-   "basket_touch_s": 0.0
+   "ball_height_m": 1.32,
+   "basket_touch_s": 0.135
   },
-  "time_s": 2.92
+  "time_s": 4.47
  },
  "recorded_result": {
   "verifier_reward_json": {
-   "success": 0,
-   "n_actions": 292,
-   "progress": 0.0,
-   "stages_reached": 0,
+   "success": 1,
+   "n_actions": 447,
+   "progress": 1.0,
+   "stages_reached": 2,
    "stages_total": 2,
-   "metric_ball_height_m": 3.073,
-   "metric_basket_touch_s": 0.0
-  }
+   "metric_ball_height_m": 1.32,
+   "metric_basket_touch_s": 0.135
+  },
+  "pool": "counted success (state success, reward 1.0, within the token budget)"
  },
  "replay": {
-  "task_dir": "roboolympics-fun-dunk-realgrasp-i00-privileged",
-  "trajectory": "jobs/codex/gpt6_astra/high/codex-gpt6_astra-high-roboolympics-fun-dunk-realgrasp-i00-privileged/roboolympics-dunk-realgrasp-i00__ZWR3UwT/artifacts/app/output/trajectory.npz",
-  "n_actions": 292,
-  "control_steps_replayed": 292,
+  "task_dir": "embodied-olympics-fun-dunk-realgrasp-i00-privileged",
+  "trajectory": "jobs/codex/gpt6_astra/high/codex-gpt6_astra-high-embodied-olympics-fun-dunk-realgrasp-i00-privileged/roboolympics-fun-dunk-realgrasp__BNuT7HK/artifacts/app/output/trajectory.npz",
+  "n_actions": 447,
+  "control_steps_replayed": 447,
   "control_hz": 100.0,
-  "every_control_steps": 1,
-  "final_state_sha": "ad8cf0efa21c118372d37a33acc3b2cdc11848f220203d49c9d898f82877dcd1",
+  "every_control_steps": 2,
+  "final_state_sha": "e6b841248f77092d0e48e8fa2fced9fc06e60cfb90472622ab589350a9d5b29a",
   "replays_agree": true,
   "mujoco": "3.3.7",
-  "original_verdict": "counted failure (verifier reward 0)"
+  "original_verdict": "counted success (verifier reward 1)"
+ },
+ "variant": {
+  "name": "dunk-realgrasp on the ring lowered to the FIBA Mini-Basketball height",
+  "ring_height_m": 2.6,
+  "board_bottom_m": 2.45,
+  "start_to_ring_m": 0.75,
+  "ran_as": "embodied-olympics-fun-dunk-realgrasp-lower-i00-privileged (formerly roboolympics-fun-dunk-realgrasp-lower; exam #412, reverted #413; deploy-only in the pool)",
+  "repo_task": "the exam's embodied-olympics-fun-dunk-realgrasp-i00-privileged is this version since PR #448 (squash 3ac12a3e86, 2026-10-19): same instance params and physics, only the family name and doc strings differ; the pool renamed the trial to that task dir (board 0203)"
  },
  "visual_changes": [
   "daylight sky and a plain ground plane in the venue's ground colour instead of the venue's painted sky, hills, trees, grandstands and perimeter boards (drawn-only scenery)",
   "the ball drawn as a basketball (orange, black channels)"
  ],
  "camera": {
+  "key_add": [
+   "hoop_rim"
+  ],
   "mode": "director",
-  "az_deg": -50.0,
+  "az_deg": -90.0,
   "az_rule": "facing-50",
   "el_deg": 22.0,
   "fovy_deg": 34.0,
   "fill": 0.72,
   "distance_m": [
-   4.205,
-   6.348
+   6.26,
+   6.406
   ],
   "robot_root": "pelvis",
-  "travel_m": 0.109,
+  "travel_m": 0.745,
   "focus": {
-   "robot+object": 190,
-   "robot": 103
+   "robot": 86,
+   "robot+object": 139
   },
   "objects": {
-   "ball": 0.48
-  }
+   "ball": 1.69
+  },
+  "note": "azimuth turned 40 degrees from the default rule (facing-50 -> facing-90, a side view from the court) and the ring kept in frame (--key-add hoop_rim): the default camera looked from behind the backboard and cut the ring off at the top, so the dunk was not visible"
  },
  "orbit_deg": 10.0,
  "holds": {
@@ -79,16 +93,17 @@
   "end_s": 1.5,
   "note": "first/last frame held, no extra simulation"
  },
- "seconds_per_frame": 3.13,
+ "seconds_per_frame": 3.17,
  "image": "ghcr.io/mll-lab-nu/efe:robofunolympics-0.2.1 (simulator) + ghcr.io/mll-lab-nu/efe:roboolympics-fun-files-0.1.0@sha256:49a33342a72b57b56742cc754e427205a30f39dfcd69a99d1d7275c0eae4a619 (rcb_rfo.py, rfo/)",
- "task_dir": "roboolympics-fun-dunk-realgrasp-i00-privileged",
+ "task_dir": "embodied-olympics-fun-dunk-realgrasp-i00-privileged",
  "agent": "codex + GPT-6 Astra (high)",
  "run": "codex-gpt6_astra-high",
  "mode": "privileged",
- "trial": "codex-gpt6_astra-high-roboolympics-fun-dunk-realgrasp-i00-privileged/roboolympics-dunk-realgrasp-i00__ZWR3UwT",
- "trial_choice": "no counted success in any run; GPT-6 Astra's privileged attempt with the most progress",
- "poster": "last frame (end of the failed attempt)",
- "host": "vast.ai RTX 4090 gpu2",
+ "trial": "codex-gpt6_astra-high-embodied-olympics-fun-dunk-realgrasp-i00-privileged/roboolympics-fun-dunk-realgrasp__BNuT7HK",
+ "trial_choice": "a counted GPT-6 Astra success in privileged mode (the only success on this task in any run)",
+ "poster": "frame 152 of the video's own frames: the ball inside the ring with both hands on it (the last frame shows the robot standing after the landing, the ball away)",
+ "host": "vast.ai RTX 4090 x12 (2 boxes, frames split round-robin)",
  "encode": "libx264 preset slow CRF 16, yuv420p, bt709, faststart",
- "created": "2026-10-19T06:14:58-0500"
+ "created": "2026-10-19T14:44:12-0500",
+ "replaces": "the earlier scene-orbit video of GPT-6 Astra's failed privileged attempt on the 3.05 m ring (ZWR3UwT)"
 }

masters/roboolympics-fun/dunk_realgrasp/video.mp4

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5ca48ccd789afd1f88b9b3eeeeaaeab135c13948dbc1c83cee432a2aff092320
-size 6537624
+oid sha256:607a2e6be2ba81792bdd0aa9a092f7d70a1d0f241d9c1c7687c3c6744eb64303
+size 5252182