blog: HQ case videos (2)

blog/humanoidbench-loco-hurdle-i00-standard--h8pcbpw/meta.json

@@ -0,0 +1,156 @@
+{
+ "slug": "humanoidbench-loco-hurdle-i00-standard--h8pcbpw",
+ "status": "rendered",
+ "task": "humanoidbench-loco-hurdle-i00-standard",
+ "bench": "humanoidbench",
+ "mode": "standard",
+ "run": "codex-gpt6_astra-high",
+ "outcome": "failed",
+ "card": "HumanoidBench: one fall ends it HumanoidBench · Hurdle GPT-6 Astra · standard · failed A running gait reached the first hurdle every time, but every jump tripped or toppled; it used all 50 resets, and the last episode fell at step 227.",
+ "source_trial": "codex/gpt6_astra/high/codex-gpt6_astra-high-humanoidbench-loco-hurdle-i00-standard/humanoidbench-loco-hurdle-i00-st__H8pCbpw",
+ "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 LTS Cycles path tracing (GPU OptiX, RTX 4090), adaptive sampling (threshold 0.01, 64-128 spp), OpenImageDenoise with albedo + normal passes; kinematic playback of the exact replay poses (the simulator's own compiled meshes, primitives and colours); daylight: a sun lamp (2 degree disc) and a Nishita sky (no sun disc), the scene's own floor colour on a large ground plane with a faint world-space variation; AgX view transform, Medium High Contrast look; camera director with a slow eased 10.0-degree orbit; 2560x1440, 30 fps, no text or overlays",
+ "rendering": "path-tracing",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "frames_rendered": 228,
+ "render_seconds": 749.1,
+ "seconds_per_frame": 3.29,
+ "time_range": {
+  "episode": "51 of 51 (the graded last episode, after reset 50; journal rows 11607-11833)",
+  "from_step": 0,
+  "to_step": 227,
+  "sim_t_from_s": 0.0,
+  "sim_t_to_s": 5.54,
+  "why": "the graded last episode, whole: the running approach (4.5 s), the jump at the first hurdle (x = 6.9 m, step ~150-165) and the fall that ends the episode at step 227, plus 1 s of the robot coming to rest",
+  "post_hold": {
+   "steps": 50,
+   "from_frame": 228,
+   "sim_t_s": 1.0,
+   "what": "after the recorded rows end (verdict and hash taken before): the last recorded row held, physics unchanged; NOT part of the trial, only the robot coming to rest. Render frames < from_frame to show the trial alone."
+  }
+ },
+ "speed": 0.6,
+ "holds": {
+  "start_s": 0.5,
+  "end_s": 1.5,
+  "note": "first/last frame held, no extra simulation"
+ },
+ "replay_check": {
+  "reproduced": true,
+  "how": "hb: every recorded row replayed from the frozen instance with the task's own tooling (physics unchanged), the verdict + final state hash taken where the episode ends; a second independent replay (the tooling's own replay()) must agree; then compared with the trial's records",
+  "recorded": {
+   "success": false,
+   "state_sha": "0ab1b950cfe0",
+   "steps": 227
+  },
+  "replayed": {
+   "success": false,
+   "state_sha": "0ab1b950cfe0",
+   "steps": 227,
+   "episode_return": 81.822,
+   "success_bar": 700.0,
+   "ended": true
+  },
+  "plain_replay": {
+   "success": false,
+   "state_sha": "0ab1b950cfe0",
+   "steps": 227,
+   "episode_return": 81.822
+  },
+  "matches": {
+   "success": true,
+   "state_sha": true,
+   "steps": true
+  },
+  "replays_agree": true,
+  "instance": {
+   "start_state_sha": "e3dad968a71f",
+   "instance_sha": "f86d05d210aa",
+   "instance_sha_file": "f86d05d210aa"
+  }
+ },
+ "camera": {
+  "mode": "director",
+  "az_deg": -69.07,
+  "az_rule": "robot-travel-70",
+  "el_deg": 22.0,
+  "fovy_deg": 34.0,
+  "fill": 0.72,
+  "distance_m": [
+   3.972,
+   6.079
+  ],
+  "robot_root": "pelvis",
+  "travel_m": 8.58,
+  "focus": {
+   "robot": 128,
+   "robot+extra": 150
+  },
+  "objects": {}
+ },
+ "visual_changes": [],
+ "export": {
+  "kind": "hb",
+  "n_actions": 227,
+  "control_steps_replayed": 227,
+  "control_hz": 50.0,
+  "every": 1,
+  "frames": 278,
+  "mujoco": "3.3.7",
+  "n_geoms": 471
+ },
+ "render": {
+  "render_s": 749.1,
+  "frames": 228,
+  "device": "GPU OptiX",
+  "look": "outdoor",
+  "min_spp": 64,
+  "max_spp": 128,
+  "w": 2560,
+  "h": 1440,
+  "blender": "4.2.23 LTS",
+  "orbit_deg": 10.0
+ },
+ "encode": "master: libx264 preset slow CRF 16, yuv420p, bt709, 30 fps, +faststart, no audio",
+ "rendered_frames_note": "only export frames 0-227 (the recorded rows of the graded last episode, up to the fall at step 227) were rendered; the exporter post-fall hold (frames 228+, not part of the trial) was left out; the end is the usual 1.5 s hold of the last frame",
+ "files": {
+  "video.mp4": {
+   "codec": "h264",
+   "profile": "High",
+   "pix_fmt": "yuv420p",
+   "width": 2560,
+   "height": 1440,
+   "fps": "30/1",
+   "frames": 288,
+   "duration_s": 9.6,
+   "bytes": 5733030,
+   "audio": false,
+   "faststart": true,
+   "sha256": "1f5e2707ebb5b07b91b3166003d5772f3101c4fca3d10d7605a81e057d425847",
+   "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": 288,
+   "duration_s": 9.6,
+   "bytes": 1866219,
+   "audio": false,
+   "faststart": true,
+   "crf": 20,
+   "sha256": "7d5bdce770c0740270f3dfdc88e472e6c9a197e800e3f61f276a141f73a4c1c1"
+  },
+  "poster.jpg": {
+   "bytes": 32306,
+   "size": "1280x720",
+   "from": "master frame at t=9.55 s",
+   "sha256": "1c2c6e2cb7cb35b3987284e3b7ff8db09655b3b0097b8130ec07a491ea32487c"
+  }
+ }
+}

blog/humanoidbench-loco-hurdle-i00-standard--h8pcbpw/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7d5bdce770c0740270f3dfdc88e472e6c9a197e800e3f61f276a141f73a4c1c1
+size 1866219

blog/myosuite-dex-hammer-use-i00-privileged--3tyvler/meta.json

@@ -0,0 +1,147 @@
+{
+ "slug": "myosuite-dex-hammer-use-i00-privileged--3tyvler",
+ "status": "rendered",
+ "task": "myosuite-dex-hammer-use-i00-privileged",
+ "bench": "myosuite",
+ "mode": "privileged",
+ "run": "codex-gpt6_astra-high",
+ "outcome": "failed",
+ "card": "MyoSuite: 45 muscles MyoSuite · Hammer use GPT-6 Astra · privileged · failed It optimized 45 muscle commands against the simulator for 90 minutes; only at minute 72 did a candidate keep the hammer in the hand all the way, and the final answer met the reference in 40 of 99 steps, up to 0.187 m off.",
+ "source_trial": "codex/gpt6_astra/high/codex-gpt6_astra-high-myosuite-dex-hammer-use-i00-privileged/myosuite-dex-hammer-use-i00-priv__3TYvLer",
+ "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 LTS Cycles path tracing (GPU OptiX, RTX 4090), adaptive sampling (threshold 0.01, 64-128 spp), OpenImageDenoise with albedo + normal passes; kinematic playback of the exact replay poses (the simulator's own compiled meshes, primitives and colours); studio: a seamless cyclorama floor in the scene's floor colour, one large soft key light, an overhead softbox, a rim light and a dim sky dome (Poly Haven studio_small_09 HDRI, CC0); AgX view transform, Medium High Contrast look; camera director with a slow eased 10.0-degree orbit; 2560x1440, 30 fps, no text or overlays",
+ "rendering": "path-tracing",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "frames_rendered": 294,
+ "render_seconds": 1379.1,
+ "seconds_per_frame": 4.69,
+ "time_range": {
+  "from_step": 0,
+  "to_step": 99,
+  "sim_t_from_s": 0.0,
+  "sim_t_to_s": 1.98,
+  "why": "the handed-in answer is only 99 control steps (1.98 s): shown whole, slowed to 0.24x (one frame every 4 physics substeps = 8 ms) so the grasp, lift and swing read; 0.5 s hold on the start pose, 1 s hold on the final pose"
+ },
+ "speed": 0.24,
+ "holds": {
+  "start_s": 0.5,
+  "end_s": 1.5,
+  "note": "first/last frame held, no extra simulation"
+ },
+ "replay_check": {
+  "reproduced": true,
+  "how": "the tooling's Session.replay() on the frozen instance (verifier path) with a read-only substep pose hook, then the verifier's own `rcb_myosuite.py replay` in a fresh process; both compared with each other and with the recorded verifier/grade.log + reward.json",
+  "recorded": {
+   "state_sha": "8cde2753231bcd8b",
+   "success": false,
+   "reason": "failure: on track at 40 of 99 steps, 0.187 m off at worst",
+   "steps": 99
+  },
+  "replayed": {
+   "live": {
+    "success": false,
+    "reason": "failure: on track at 40 of 99 steps, 0.187 m off at worst",
+    "steps": 99,
+    "state_sha": "8cde2753231bcd8b",
+    "track_ok": 40,
+    "instance_exact": true
+   },
+   "plain_replay": {
+    "success": false,
+    "reason": "failure: on track at 40 of 99 steps, 0.187 m off at worst",
+    "steps": 99,
+    "state_sha": "8cde2753231bcd8b",
+    "instance_ok": true
+   }
+  },
+  "replays_agree": true
+ },
+ "camera": {
+  "mode": "director",
+  "az_deg": -170.0,
+  "el_deg": 18.0,
+  "fovy_deg": 30.0,
+  "fill": 0.72,
+  "distance_m": [
+   0.972,
+   1.067
+  ],
+  "smooth_frames": 24.0,
+  "target": [
+   0.0881,
+   -0.1295,
+   0.0957
+  ],
+  "keeps_in_frame": "hand + forearm bones, hammer (upper arm/torso and table may crop)"
+ },
+ "visual_changes": [
+  "muscle-tendon paths (MuJoCo tendons, the 39 muscles) drawn as per-frame cylinder segments + joint spheres in the model's tendon colour (0.95,0.3,0.3); radius 1.5 mm (model tendon_width 1 mm, x1.5 for legibility); colour does not follow activation",
+  "reference target: the model's site 'target' (sphere r 3 cm, rgba 0.1,0.8,0.2,0.4) that MyoDex moves to the reference object position every control step; hidden until MyoDex first sets it (frame 0 / start hold: it sits at its unset default inside the table)",
+  "not drawn (MuJoCo's default hidden groups / invisible geoms): wrap geoms (group 3), skin capsules + duplicate collision table (group 4), alpha-0 'full_body' and 'hammer_visual' meshes",
+  "MuJoCo ground planes replaced by the renderer's floor at the main plane's height (z -0.76, colour 0.1 grey); table and bones in flat colours (material x texture mean)",
+  "upper arm / shoulder bones are drawn but not kept in frame by the camera (the hand, forearm and hammer are)"
+ ],
+ "export": {
+  "kind": "myosuite",
+  "n_actions": 99,
+  "control_steps_replayed": 99,
+  "control_hz": 50.0,
+  "every": 0.4,
+  "frames": 294,
+  "mujoco": "3.6.0",
+  "n_geoms": 622
+ },
+ "render": {
+  "render_s": 1379.1,
+  "frames": 294,
+  "device": "GPU OptiX",
+  "look": "studio",
+  "min_spp": 64,
+  "max_spp": 128,
+  "w": 2560,
+  "h": 1440,
+  "blender": "4.2.23 LTS",
+  "orbit_deg": 10.0
+ },
+ "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": 354,
+   "duration_s": 11.8,
+   "bytes": 6146719,
+   "audio": false,
+   "faststart": true,
+   "sha256": "436be929a494919cc8bd9f04f69394de23235d0a34ad7bede0cdaadb7be96037",
+   "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": 354,
+   "duration_s": 11.8,
+   "bytes": 1929573,
+   "audio": false,
+   "faststart": true,
+   "crf": 20,
+   "sha256": "888168bb07933238f217fcd1fb1a080b1d57497551dd0865a95ae5709b5257ba"
+  },
+  "poster.jpg": {
+   "bytes": 27813,
+   "size": "1280x720",
+   "from": "master frame at t=11.75 s",
+   "sha256": "791ef500c607c1f948c3838657d7573773024f848011b8794dfaed94ffdddfb5"
+  }
+ }
+}

blog/myosuite-dex-hammer-use-i00-privileged--3tyvler/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:888168bb07933238f217fcd1fb1a080b1d57497551dd0865a95ae5709b5257ba
+size 1929573