blog: HQ case videos (1)

blog/robowits-retrieve-cube-i00-privileged--4idoq4k/meta.json

@@ -0,0 +1,163 @@
+{
+ "slug": "robowits-retrieve-cube-i00-privileged--4idoq4k",
+ "status": "rendered",
+ "task": "robowits-retrieve-cube-i00-privileged",
+ "bench": "robowits",
+ "mode": "privileged",
+ "run": "claude_code-opus_5_5-high",
+ "outcome": "success",
+ "card": "Claude Opus 5.5, privileged, solved",
+ "source_trial": "claude_code/opus_5_5/high/claude_code-opus_5_5-high-robowits-retrieve-cube-i00-privileged/robowits-retrieve-cube-i00-privi__4idoq4K",
+ "what": "re-rendered: kinematic playback of this trial's own replay (its recorded actions replayed from the task's frozen instance with the task's own tooling, physics unchanged) in Blender Cycles",
+ "renderer": "Blender 4.2.23 LTS Cycles path tracing (GPU OptiX, RTX 4090), adaptive sampling (threshold 0.01, 32-96 spp), OpenImageDenoise with albedo + normal passes; kinematic playback of the exact replay poses (the simulator's own visual meshes and colours); the suite's HF master look: its static camera, Poly Haven studio_small_09 HDRI (CC0) x0.7 + one warm directional light (colour 1.0,0.92,0.80), floor 0.24,0.24,0.25, table 0.52,0.38,0.25; AgX, Medium High Contrast, exposure -0.3; 2560x1440, 30 fps, no text or overlays",
+ "rendering": "path-tracing",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "frames_rendered": 496,
+ "render_seconds": 4118.1,
+ "seconds_per_frame": 8.3,
+ "time_range": {
+  "from_step": 0,
+  "to_step": 990,
+  "sim_t_from_s": 0.0,
+  "sim_t_to_s": 33.0,
+  "why": "the whole graded replay: the trial's 900 submitted action rows from the frozen instance, then the verifier's 90-step settle/check window (3 s) in which the cube must stay on the target, released and at rest"
+ },
+ "speed": 2.0,
+ "holds": {
+  "start_s": 0.5,
+  "end_s": 1.5,
+  "note": "first/last frame held, no extra simulation"
+ },
+ "replay_check": {
+  "reproduced": true,
+  "how": "the submitted artifacts/app/output/trajectory.npz (900 x 16 JOINT_ABS rows, sha256 f2f8ec5c48f2bef4...) replayed exactly as the verifier's grade.py does it: rcb_robowits.Session(retrieve_cube, tests/instance.json) on Genesis 0.4.7's CPU backend in the task's own image, Session.replay() incl. the settle window; every visual geom's world pose recorded after each control step (kinematic read-out, physics untouched); a second plain replay in the same process must agree",
+  "recorded": {
+   "success": 1,
+   "all_success": 1,
+   "deterministic": 1,
+   "sim_steps": 990,
+   "goal_at_end": 1,
+   "final_reward": 1.0,
+   "settle_max_move_mm": 0.24,
+   "trial_host_instance_sha": "df0f5785c0335b3a",
+   "trial_host_final_state_sha": "af3d35d93aa3547a (agent's own grader-equivalent replay)",
+   "cube_final_xy_m": [
+    0.448,
+    0.189
+   ]
+  },
+  "replayed": {
+   "success": true,
+   "sim_steps": 990,
+   "goal_at_end": true,
+   "progress": 1.0,
+   "settle": {
+    "success": true,
+    "why": "the goal held, released and at rest, through the check window",
+    "steps": 90,
+    "max_move_m": 0.00042
+   },
+   "state_sha": "a91c503b5c246b40",
+   "instance_sha": "8b96edf97300e86f",
+   "cube_final_xy_m": [
+    0.4287,
+    0.1512
+   ]
+  },
+  "plain_replay": {
+   "success": true,
+   "state_sha": "a91c503b5c246b40",
+   "sim_steps": 990
+  },
+  "replays_agree": true,
+  "matches": {
+   "success": true,
+   "sim_steps": true,
+   "goal_at_end": true,
+   "state_sha": false
+  },
+  "note": "same verdict (success, 990 steps, goal at the end of the rows, settled), but not bit-identical: the frozen instance rebuilds with different last bits on every CPU model (instance sha 0567a966 frozen, df0f5785 on the trial's host, 8b96edf9 on the render host; poses within the tooling's tolerance on all), and the cube's slide out of the tipped container amplifies them: here it lands on the target about 4 cm from where it landed in the trial ((0.429, 0.151) vs (0.448, 0.189) m). A third CPU model (Zen 2) rebuilt the instance as a728233a and the same rows FAILED there (cube not on the target, progress 0.696), so this trajectory's success is host-sensitive; the trial's host is no longer running. The render shows the replay on the render host, which passes like the graded one."
+ },
+ "camera": {
+  "mode": "static",
+  "source": "the suite master's camera (masters/robowits/retrieve_cube meta.json: Genesis pos/lookat, vertical fov 28 deg)",
+  "pos": [
+   1.6516698021629805,
+   -0.7005306236236384,
+   1.771520501721366
+  ],
+  "lookat": [
+   0.4269010400646318,
+   0.006589950868887917,
+   0.853101560958101
+  ],
+  "fov": 28.0
+ },
+ "visual_changes": [
+  "floor and table recoloured as in the suite master (floor 0.24,0.24,0.25; table 0.52,0.38,0.25); the simulator's checker ground plane replaced by a plain floor",
+  "no text or overlays"
+ ],
+ "export": {
+  "kind": "robowits",
+  "n_actions": 900,
+  "control_steps_replayed": 990,
+  "control_hz": 30,
+  "every": 2,
+  "frames": 496
+ },
+ "render": {
+  "renderer": "Blender 4.2.23 LTS Cycles path tracing (GPU OptiX, RTX 4090), adaptive sampling (threshold 0.01, 32-96 spp), OpenImageDenoise with albedo + normal passes; kinematic playback of the exact replay poses (the simulator's own visual meshes and colours)",
+  "look": "the suite's HF master look: its static camera, Poly Haven studio_small_09 HDRI (CC0) x0.7 + one warm directional light (colour 1.0,0.92,0.80), floor 0.24,0.24,0.25, table 0.52,0.38,0.25; AgX, Medium High Contrast, exposure -0.3",
+  "frames": 496,
+  "render_s": 4118.1,
+  "w": 2560,
+  "h": 1440,
+  "blender": "4.2.23 LTS",
+  "device": "GPU OptiX",
+  "min_spp": 32,
+  "max_spp": 96,
+  "gpus": "8 RTX 4090 runs (frames split across GPUs)",
+  "every": 2
+ },
+ "encode": "master: libx264 preset slow CRF 16, yuv420p, bt709, 30 fps, +faststart, no audio",
+ "files": {
+  "video.mp4": {
+   "codec": "h264",
+   "profile": "High",
+   "pix_fmt": "yuv420p",
+   "width": 2560,
+   "height": 1440,
+   "fps": "30/1",
+   "frames": 556,
+   "duration_s": 18.534,
+   "bytes": 5962596,
+   "audio": false,
+   "faststart": true,
+   "sha256": "ee30a4639de7aba00762057859d77bdb8471c40f7bc92de0ad6351f0812fab73",
+   "note": "stream copy of the master (remuxed with +faststart)"
+  },
+  "web.mp4": {
+   "codec": "h264",
+   "profile": "High",
+   "pix_fmt": "yuv420p",
+   "width": 1920,
+   "height": 1080,
+   "fps": "30/1",
+   "frames": 556,
+   "duration_s": 18.534,
+   "bytes": 1728700,
+   "audio": false,
+   "faststart": true,
+   "crf": 20,
+   "sha256": "165d8e18e7e61fd07bcd6b2eaf69306f13d6471cc84d39e989aea702b9a51f3d"
+  },
+  "poster.jpg": {
+   "bytes": 28192,
+   "size": "1280x720",
+   "from": "master frame at t=18.48 s",
+   "sha256": "59ac997136b600fb680ad5b4a1043cdbd0b1c934f3e6e73d918dce74ac44b352"
+  }
+ }
+}

blog/robowits-retrieve-cube-i00-privileged--4idoq4k/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:165d8e18e7e61fd07bcd6b2eaf69306f13d6471cc84d39e989aea702b9a51f3d
+size 1728700