EmbodiedSWE masters: 24 tasks (2K path-traced), highlights, teaser

masters/embodiedswe/assembly_allen_bolt_franka_osc/meta.json

@@ -0,0 +1,52 @@
+{
+ "suite": "embodiedswe",
+ "task": "assembly_allen_bolt_franka_osc",
+ "task_dir": "embodiedswe-assembly-allen-bolt-franka-osc-i00-standard",
+ "source": "agent-success:codex-gpt6_luna-xhigh-embodiedswe-assembly-allen-bolt-franka-osc-i00-standard/embodiedswe-assembly-allen-bolt__AaZsDRG",
+ "agent": "Codex + GPT-6 Luna (xhigh)",
+ "run": "codex-gpt6_luna-xhigh",
+ "mode": "standard",
+ "trial": "AaZsDRG",
+ "trial_choice": "the only run that solved this task (GPT-6 Luna, counted standard success); the other runs did not attempt it (screening: they run only tasks Luna fails in every mode)",
+ "renderer": "isaac-rtx PathTracing spp=32 totalSpp=32 maxBounces=6 optix-denoiser firefly-filter (Isaac Sim 5.1 / Isaac Lab 2.3.2; dedicated render camera at the scene's own viewpoint, rendered natively at 2560x1440, focal 18.0 mm; recorded states posed, render only)",
+ "rendering": "path-tracing",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "frames": 786,
+ "seconds": 26.2,
+ "speed": 32.0,
+ "capture": "one recorded state every 16 control steps (1.067 s of simulated time) per video frame = 32x real time, 726 rendered frames; 0.5 s hold of the first frame, 1.5 s hold of the final frame",
+ "camera": {
+  "eye": [
+   1.1,
+   -0.72,
+   0.85
+  ],
+  "target": [
+   0.36,
+   0.05,
+   0.2
+  ],
+  "focal_mm": 18.0,
+  "room_camera": [
+   [
+    1.1,
+    -0.72,
+    0.85
+   ],
+   [
+    0.36,
+    0.05,
+    0.2
+   ]
+  ]
+ },
+ "success_check": true,
+ "success_check_how": "the video poses the graded episode's own recorded scene states (trajectory.npz frame_state: every asset's root pose and velocity and joint state, recorded during the graded live standard episode by the simulator service) on the task's own scene build (InteractiveScene.reset_to, no physics step), so it shows the graded run itself. Start state = the frozen instance (instance_ok 1, NVIDIA L40S). The task's success predicate on the posed final state: 1. Verdict (reward.json): success 1, progress 1.0.",
+ "image": "efe-private:embodiedswe-0.3.0 (sim service: efe:embodiedswe-standard-files-0.2.0)",
+ "core": false,
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, +faststart",
+ "poster": "the final frame",
+ "created": "2026-10-30"
+}

masters/embodiedswe/assembly_allen_bolt_franka_osc/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1a88e216eef9232b15c4c6a9c35d6129ef0c6068a12fd2e59f8a86da02da3399
+size 16549848

masters/embodiedswe/assembly_bulb_franka_osc/meta.json

@@ -0,0 +1,52 @@
+{
+ "suite": "embodiedswe",
+ "task": "assembly_bulb_franka_osc",
+ "task_dir": "embodiedswe-assembly-bulb-franka-osc-i00-privileged",
+ "source": "agent-success:codex-gpt6_astra-high-embodiedswe-assembly-bulb-franka-osc-i00-privileged/embodiedswe-assembly-bulb-franka__KWguxBe",
+ "agent": "Codex + GPT-6 Astra (high)",
+ "run": "codex-gpt6_astra-high",
+ "mode": "privileged",
+ "trial": "KWguxBe",
+ "trial_choice": "counted privileged success; order: GPT-6 Astra > Claude Opus 5.5 > Gemini 4 Argon > Gemini 3.8 Flash, then the task's default mode (standard) before privileged",
+ "renderer": "isaac-rtx PathTracing spp=32 totalSpp=32 maxBounces=6 optix-denoiser firefly-filter (Isaac Sim 5.1 / Isaac Lab 2.3.2; dedicated render camera at the scene's own viewpoint, rendered natively at 2560x1440, focal 18.0 mm; recorded states posed, render only)",
+ "rendering": "path-tracing",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "frames": 623,
+ "seconds": 20.77,
+ "speed": 8.0,
+ "capture": "one recorded state every 4 control steps (0.267 s of simulated time) per video frame = 8x real time, 563 rendered frames; 0.5 s hold of the first frame, 1.5 s hold of the final frame",
+ "camera": {
+  "eye": [
+   1.1,
+   -0.72,
+   0.85
+  ],
+  "target": [
+   0.36,
+   0.05,
+   0.2
+  ],
+  "focal_mm": 18.0,
+  "room_camera": [
+   [
+    1.1,
+    -0.72,
+    0.85
+   ],
+   [
+    0.36,
+    0.05,
+    0.2
+   ]
+  ]
+ },
+ "success_check": true,
+ "success_check_how": "the video poses the graded episode's own recorded scene states (trajectory.npz frame_state: every asset's root pose and velocity and joint state, recorded during the graded verifier rollout) on the task's own scene build (InteractiveScene.reset_to, no physics step), so it shows the graded run itself. Start state = the frozen instance (instance_ok 1, NVIDIA L40S). The task's success predicate on the posed final state: 1. Verdict (reward.json): success 1, progress 1.0. The verifier's own camera replay of the actions ended on the graded final state: camera_replay_match 1.",
+ "image": "efe-private:embodiedswe-0.3.0 (verifier tooling: efe:embodiedswe-files-0.3.1)",
+ "core": true,
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, +faststart",
+ "poster": "the final frame",
+ "created": "2026-10-30"
+}

masters/embodiedswe/assembly_bulb_franka_osc/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7d7bc37b521100971112c4fdf8676cfcaf730aea90cd9ed80ad8445e908bc118
+size 8349546

masters/embodiedswe/assembly_ikea_table_bimanual_franka_osc/meta.json

@@ -0,0 +1,54 @@
+{
+ "suite": "embodiedswe",
+ "task": "assembly_ikea_table_bimanual_franka_osc",
+ "task_dir": "embodiedswe-assembly-ikea-table-bimanual-franka-osc-i00-privileged",
+ "source": "scene-orbit",
+ "agent": "Codex + GPT-6 Astra (high)",
+ "run": "codex-gpt6_astra-high",
+ "mode": "privileged",
+ "trial": "464xuPn",
+ "trial_choice": "no evaluated run has solved this task yet; this is the failed attempt with the most progress (Codex + GPT-6 Astra (high), privileged, progress 0.96); not a solved run",
+ "renderer": "isaac-rtx PathTracing spp=32 totalSpp=32 maxBounces=6 optix-denoiser firefly-filter (Isaac Sim 5.1 / Isaac Lab 2.3.2; dedicated render camera at the scene's own viewpoint, rendered natively at 2560x1440, focal 18.0 mm; recorded states posed, render only)",
+ "rendering": "path-tracing",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "frames": 939,
+ "seconds": 31.3,
+ "speed": 16.0,
+ "capture": "one recorded state every 8 control steps (0.533 s of simulated time) per video frame = 16x real time, 879 rendered frames; 0.5 s hold of the first frame, 1.5 s hold of the final frame",
+ "camera": {
+  "eye": [
+   -0.35,
+   -1.9,
+   1.75
+  ],
+  "target": [
+   -0.4,
+   0.0,
+   1.05
+  ],
+  "focal_mm": 18.0,
+  "room_camera": [
+   [
+    -0.35,
+    -1.9,
+    1.75
+   ],
+   [
+    -0.4,
+    0.0,
+    1.05
+   ]
+  ]
+ },
+ "success_check": false,
+ "success_check_how": "the video poses the graded episode's own recorded scene states (trajectory.npz frame_state: every asset's root pose and velocity and joint state, recorded during the graded verifier rollout) on the task's own scene build (InteractiveScene.reset_to, no physics step), so it shows the graded run itself. Start state = the frozen instance (instance_ok 1, NVIDIA L40S). The task's success predicate on the posed final state: 0. Verdict (reward.json): success 0, progress 0.9586. The verifier's own camera replay of the actions ended on the graded final state: camera_replay_match 0.",
+ "image": "efe-private:embodiedswe-0.3.0 (verifier tooling: efe:embodiedswe-files-0.3.1)",
+ "core": true,
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, +faststart",
+ "poster": "the final frame",
+ "created": "2026-10-30",
+ "why_no_video": "No evaluated run has solved this task yet. The video shows the failed attempt with the most progress (Codex + GPT-6 Astra (high), privileged, progress 0.96), not a solved run.",
+ "shows": "agent-attempt (failed):codex-gpt6_astra-high/464xuPn"
+}

masters/embodiedswe/assembly_ikea_table_bimanual_franka_osc/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b0fb71cc5a0fef286076bc6477046b431a7788406034e0d45efc1c1651116bcc
+size 14786984

masters/embodiedswe/assembly_nut_thread_franka_osc/meta.json

@@ -0,0 +1,52 @@
+{
+ "suite": "embodiedswe",
+ "task": "assembly_nut_thread_franka_osc",
+ "task_dir": "embodiedswe-assembly-nut-thread-franka-osc-i00-standard",
+ "source": "agent-success:codex-gpt6_luna-xhigh-embodiedswe-assembly-nut-thread-franka-osc-i00-standard/embodiedswe-assembly-nut-thread__Vw6ueVC",
+ "agent": "Codex + GPT-6 Luna (xhigh)",
+ "run": "codex-gpt6_luna-xhigh",
+ "mode": "standard",
+ "trial": "Vw6ueVC",
+ "trial_choice": "the only run that solved this task (GPT-6 Luna, counted standard success); the other runs did not attempt it (screening: they run only tasks Luna fails in every mode)",
+ "renderer": "isaac-rtx PathTracing spp=32 totalSpp=32 maxBounces=6 optix-denoiser firefly-filter (Isaac Sim 5.1 / Isaac Lab 2.3.2; dedicated render camera at the scene's own viewpoint, rendered natively at 2560x1440, focal 18.0 mm; recorded states posed, render only)",
+ "rendering": "path-tracing",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "frames": 834,
+ "seconds": 27.8,
+ "speed": 16.0,
+ "capture": "one recorded state every 8 control steps (0.533 s of simulated time) per video frame = 16x real time, 774 rendered frames; 0.5 s hold of the first frame, 1.5 s hold of the final frame",
+ "camera": {
+  "eye": [
+   1.1,
+   -0.72,
+   0.85
+  ],
+  "target": [
+   0.36,
+   0.05,
+   0.2
+  ],
+  "focal_mm": 18.0,
+  "room_camera": [
+   [
+    1.1,
+    -0.72,
+    0.85
+   ],
+   [
+    0.36,
+    0.05,
+    0.2
+   ]
+  ]
+ },
+ "success_check": true,
+ "success_check_how": "the video poses the graded episode's own recorded scene states (trajectory.npz frame_state: every asset's root pose and velocity and joint state, recorded during the graded live standard episode by the simulator service) on the task's own scene build (InteractiveScene.reset_to, no physics step), so it shows the graded run itself. Start state = the frozen instance (instance_ok 1, NVIDIA L40S). The task's success predicate on the posed final state: 1. Verdict (reward.json): success 1, progress 1.0.",
+ "image": "efe-private:embodiedswe-0.3.0 (sim service: efe:embodiedswe-standard-files-0.2.0)",
+ "core": false,
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, +faststart",
+ "poster": "the final frame",
+ "created": "2026-10-30"
+}

masters/embodiedswe/assembly_nut_thread_franka_osc/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:099a43130b8b47c98b050ba2dd3292692776905e1e473a451e5d00aee851b931
+size 17001493

masters/embodiedswe/assembly_pc_gpu_franka_osc/meta.json

@@ -0,0 +1,52 @@
+{
+ "suite": "embodiedswe",
+ "task": "assembly_pc_gpu_franka_osc",
+ "task_dir": "embodiedswe-assembly-pc-gpu-franka-osc-i00-standard",
+ "source": "agent-success:codex-gpt6_astra-high-embodiedswe-assembly-pc-gpu-franka-osc-i00-standard/embodiedswe-assembly-pc-gpu-fran__AoPoRzm",
+ "agent": "Codex + GPT-6 Astra (high)",
+ "run": "codex-gpt6_astra-high",
+ "mode": "standard",
+ "trial": "AoPoRzm",
+ "trial_choice": "counted standard success; order: GPT-6 Astra > Claude Opus 5.5 > Gemini 4 Argon > Gemini 3.8 Flash, then the task's default mode (standard) before privileged",
+ "renderer": "isaac-rtx PathTracing spp=32 totalSpp=32 maxBounces=6 optix-denoiser firefly-filter (Isaac Sim 5.1 / Isaac Lab 2.3.2; dedicated render camera at the scene's own viewpoint, rendered natively at 2560x1440, focal 18.0 mm; recorded states posed, render only)",
+ "rendering": "path-tracing",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "frames": 693,
+ "seconds": 23.1,
+ "speed": 24.0,
+ "capture": "one recorded state every 12 control steps (0.800 s of simulated time) per video frame = 24x real time, 633 rendered frames; 0.5 s hold of the first frame, 1.5 s hold of the final frame",
+ "camera": {
+  "eye": [
+   1.22,
+   -1.2,
+   1.08
+  ],
+  "target": [
+   0.3,
+   -0.02,
+   0.2
+  ],
+  "focal_mm": 18.0,
+  "room_camera": [
+   [
+    1.22,
+    -1.2,
+    1.08
+   ],
+   [
+    0.3,
+    -0.02,
+    0.2
+   ]
+  ]
+ },
+ "success_check": true,
+ "success_check_how": "the video poses the graded episode's own recorded scene states (trajectory.npz frame_state: every asset's root pose and velocity and joint state, recorded during the graded live standard episode by the simulator service) on the task's own scene build (InteractiveScene.reset_to, no physics step), so it shows the graded run itself. Start state = the frozen instance (instance_ok 1, NVIDIA L40S). The task's success predicate on the posed final state: 1. Verdict (reward.json): success 1, progress 1.0.",
+ "image": "efe-private:embodiedswe-0.3.0 (sim service: efe:embodiedswe-standard-files-0.2.0)",
+ "core": true,
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, +faststart",
+ "poster": "the final frame",
+ "created": "2026-10-30"
+}

masters/embodiedswe/assembly_pc_gpu_franka_osc/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7cd461aacc297eff983f8727b9096fbacdbf435ad0c4af8617ca1fd9cdd83514
+size 11711462

masters/embodiedswe/assembly_pc_gpu_ram_franka_osc/meta.json

@@ -0,0 +1,52 @@
+{
+ "suite": "embodiedswe",
+ "task": "assembly_pc_gpu_ram_franka_osc",
+ "task_dir": "embodiedswe-assembly-pc-gpu-ram-franka-osc-i00-privileged",
+ "source": "agent-success:codex-gpt6_astra-high-embodiedswe-assembly-pc-gpu-ram-franka-osc-i00-privileged/embodiedswe-assembly-pc-gpu-ram__7z5tXVi",
+ "agent": "Codex + GPT-6 Astra (high)",
+ "run": "codex-gpt6_astra-high",
+ "mode": "privileged",
+ "trial": "7z5tXVi",
+ "trial_choice": "counted privileged success; order: GPT-6 Astra > Claude Opus 5.5 > Gemini 4 Argon > Gemini 3.8 Flash, then the task's default mode (standard) before privileged",
+ "renderer": "isaac-rtx PathTracing spp=32 totalSpp=32 maxBounces=6 optix-denoiser firefly-filter (Isaac Sim 5.1 / Isaac Lab 2.3.2; dedicated render camera at the scene's own viewpoint, rendered natively at 2560x1440, focal 18.0 mm; recorded states posed, render only)",
+ "rendering": "path-tracing",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "frames": 672,
+ "seconds": 22.4,
+ "speed": 8.0,
+ "capture": "one recorded state every 4 control steps (0.267 s of simulated time) per video frame = 8x real time, 612 rendered frames; 0.5 s hold of the first frame, 1.5 s hold of the final frame",
+ "camera": {
+  "eye": [
+   1.22,
+   -1.2,
+   1.08
+  ],
+  "target": [
+   0.3,
+   -0.02,
+   0.2
+  ],
+  "focal_mm": 18.0,
+  "room_camera": [
+   [
+    1.22,
+    -1.2,
+    1.08
+   ],
+   [
+    0.3,
+    -0.02,
+    0.2
+   ]
+  ]
+ },
+ "success_check": true,
+ "success_check_how": "the video poses the graded episode's own recorded scene states (trajectory.npz frame_state: every asset's root pose and velocity and joint state, recorded during the graded verifier rollout) on the task's own scene build (InteractiveScene.reset_to, no physics step), so it shows the graded run itself. Start state = the frozen instance (instance_ok 1, NVIDIA L40S). The task's success predicate on the posed final state: 1. Verdict (reward.json): success 1, progress 1.0. The verifier's own camera replay of the actions ended on the graded final state: camera_replay_match 1.",
+ "image": "efe-private:embodiedswe-0.3.0 (verifier tooling: efe:embodiedswe-files-0.3.1)",
+ "core": true,
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, +faststart",
+ "poster": "the final frame",
+ "created": "2026-10-30"
+}

masters/embodiedswe/assembly_pc_gpu_ram_franka_osc/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cd2fbdcf181cedf4eb8c61514e9d8ecdc31fcc337ef552ed3bc7936486d34307
+size 8701633

masters/embodiedswe/assembly_pc_motherboard_franka_osc/meta.json

@@ -0,0 +1,52 @@
+{
+ "suite": "embodiedswe",
+ "task": "assembly_pc_motherboard_franka_osc",
+ "task_dir": "embodiedswe-assembly-pc-motherboard-franka-osc-i00-standard",
+ "source": "agent-success:codex-gpt6_astra-high-embodiedswe-assembly-pc-motherboard-franka-osc-i00-standard/embodiedswe-assembly-pc-motherbo__advYJ7n",
+ "agent": "Codex + GPT-6 Astra (high)",
+ "run": "codex-gpt6_astra-high",
+ "mode": "standard",
+ "trial": "advYJ7n",
+ "trial_choice": "counted standard success; order: GPT-6 Astra > Claude Opus 5.5 > Gemini 4 Argon > Gemini 3.8 Flash, then the task's default mode (standard) before privileged",
+ "renderer": "isaac-rtx PathTracing spp=32 totalSpp=32 maxBounces=6 optix-denoiser firefly-filter (Isaac Sim 5.1 / Isaac Lab 2.3.2; dedicated render camera at the scene's own viewpoint, rendered natively at 2560x1440, focal 18.0 mm; recorded states posed, render only)",
+ "rendering": "path-tracing",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "frames": 793,
+ "seconds": 26.43,
+ "speed": 48.0,
+ "capture": "one recorded state every 24 control steps (1.600 s of simulated time) per video frame = 48x real time, 733 rendered frames; 0.5 s hold of the first frame, 1.5 s hold of the final frame",
+ "camera": {
+  "eye": [
+   1.22,
+   -1.2,
+   1.08
+  ],
+  "target": [
+   0.3,
+   -0.02,
+   0.2
+  ],
+  "focal_mm": 18.0,
+  "room_camera": [
+   [
+    1.22,
+    -1.2,
+    1.08
+   ],
+   [
+    0.3,
+    -0.02,
+    0.2
+   ]
+  ]
+ },
+ "success_check": true,
+ "success_check_how": "the video poses the graded episode's own recorded scene states (trajectory.npz frame_state: every asset's root pose and velocity and joint state, recorded during the graded live standard episode by the simulator service) on the task's own scene build (InteractiveScene.reset_to, no physics step), so it shows the graded run itself. Start state = the frozen instance (instance_ok 1, NVIDIA L40S). The task's success predicate on the posed final state: 1. Verdict (reward.json): success 1, progress 1.0.",
+ "image": "efe-private:embodiedswe-0.3.0 (sim service: efe:embodiedswe-standard-files-0.2.0)",
+ "core": true,
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, +faststart",
+ "poster": "the final frame",
+ "created": "2026-10-30"
+}

masters/embodiedswe/assembly_pc_motherboard_franka_osc/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8477997aa6431a7f4f8a6aabaf043a5a606b9b3d8be40f3eb00e9a60688434e3
+size 13527641

masters/embodiedswe/assembly_pc_ram_franka_osc/meta.json

@@ -0,0 +1,52 @@
+{
+ "suite": "embodiedswe",
+ "task": "assembly_pc_ram_franka_osc",
+ "task_dir": "embodiedswe-assembly-pc-ram-franka-osc-i00-privileged",
+ "source": "agent-success:codex-gpt6_luna-xhigh-embodiedswe-assembly-pc-ram-franka-osc-i00-privileged/embodiedswe-assembly-pc-ram-fran__3KkLgig",
+ "agent": "Codex + GPT-6 Luna (xhigh)",
+ "run": "codex-gpt6_luna-xhigh",
+ "mode": "privileged",
+ "trial": "3KkLgig",
+ "trial_choice": "the only run that solved this task (GPT-6 Luna, counted privileged success); the other runs did not attempt it (screening: they run only tasks Luna fails in every mode)",
+ "renderer": "isaac-rtx PathTracing spp=32 totalSpp=32 maxBounces=6 optix-denoiser firefly-filter (Isaac Sim 5.1 / Isaac Lab 2.3.2; dedicated render camera at the scene's own viewpoint, rendered natively at 2560x1440, focal 18.0 mm; recorded states posed, render only)",
+ "rendering": "path-tracing",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "frames": 703,
+ "seconds": 23.43,
+ "speed": 4.0,
+ "capture": "one recorded state every 2 control steps (0.133 s of simulated time) per video frame = 4x real time, 643 rendered frames; 0.5 s hold of the first frame, 1.5 s hold of the final frame",
+ "camera": {
+  "eye": [
+   1.22,
+   -1.2,
+   1.08
+  ],
+  "target": [
+   0.3,
+   -0.02,
+   0.2
+  ],
+  "focal_mm": 18.0,
+  "room_camera": [
+   [
+    1.22,
+    -1.2,
+    1.08
+   ],
+   [
+    0.3,
+    -0.02,
+    0.2
+   ]
+  ]
+ },
+ "success_check": true,
+ "success_check_how": "the video poses the graded episode's own recorded scene states (trajectory.npz frame_state: every asset's root pose and velocity and joint state, recorded during the graded verifier rollout) on the task's own scene build (InteractiveScene.reset_to, no physics step), so it shows the graded run itself. Start state = the frozen instance (instance_ok 1, NVIDIA L40S). The task's success predicate on the posed final state: 1. Verdict (reward.json): success 1, progress 1.0. The verifier's own camera replay of the actions ended on the graded final state: camera_replay_match 0.",
+ "image": "efe-private:embodiedswe-0.3.0 (verifier tooling: efe:embodiedswe-files-0.3.1)",
+ "core": false,
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, +faststart",
+ "poster": "the final frame",
+ "created": "2026-10-30"
+}

masters/embodiedswe/assembly_pc_ram_franka_osc/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9739e8da18d00daa7b0bd0d406ad2929db850c0bc2889add75fc909baa494d47
+size 7974405

masters/embodiedswe/assembly_so101_bimanual_franka_osc/meta.json

@@ -0,0 +1,54 @@
+{
+ "suite": "embodiedswe",
+ "task": "assembly_so101_bimanual_franka_osc",
+ "task_dir": "embodiedswe-assembly-so101-bimanual-franka-osc-i00-privileged",
+ "source": "scene-orbit",
+ "agent": "Codex + GPT-6 Astra (high)",
+ "run": "codex-gpt6_astra-high",
+ "mode": "privileged",
+ "trial": "tzC9EQD",
+ "trial_choice": "no evaluated run has solved this task yet; this is the failed attempt with the most progress (Codex + GPT-6 Astra (high), privileged, progress 0.14); not a solved run",
+ "renderer": "isaac-rtx PathTracing spp=32 totalSpp=32 maxBounces=6 optix-denoiser firefly-filter (Isaac Sim 5.1 / Isaac Lab 2.3.2; dedicated render camera at the scene's own viewpoint, rendered natively at 2560x1440, focal 18.0 mm; recorded states posed, render only)",
+ "rendering": "path-tracing",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "frames": 482,
+ "seconds": 16.07,
+ "speed": 4.0,
+ "capture": "one recorded state every 2 control steps (0.133 s of simulated time) per video frame = 4x real time, 422 rendered frames; 0.5 s hold of the first frame, 1.5 s hold of the final frame",
+ "camera": {
+  "eye": [
+   1.15,
+   1.1,
+   1.65
+  ],
+  "target": [
+   0.4,
+   -0.1,
+   1.0
+  ],
+  "focal_mm": 18.0,
+  "room_camera": [
+   [
+    1.15,
+    1.1,
+    1.65
+   ],
+   [
+    0.4,
+    -0.1,
+    1.0
+   ]
+  ]
+ },
+ "success_check": false,
+ "success_check_how": "the video poses the graded episode's own recorded scene states (trajectory.npz frame_state: every asset's root pose and velocity and joint state, recorded during the graded verifier rollout) on the task's own scene build (InteractiveScene.reset_to, no physics step), so it shows the graded run itself. Start state = the frozen instance (instance_ok 1, NVIDIA L40S). The task's success predicate on the posed final state: 0. Verdict (reward.json): success 0, progress 0.1429. The verifier's own camera replay of the actions ended on the graded final state: camera_replay_match 0.",
+ "image": "efe-private:embodiedswe-0.3.0 (verifier tooling: efe:embodiedswe-files-0.3.1)",
+ "core": true,
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, +faststart",
+ "poster": "the final frame",
+ "created": "2026-10-30",
+ "why_no_video": "No evaluated run has solved this task yet. The video shows the failed attempt with the most progress (Codex + GPT-6 Astra (high), privileged, progress 0.14), not a solved run.",
+ "shows": "agent-attempt (failed):codex-gpt6_astra-high/tzC9EQD"
+}

masters/embodiedswe/assembly_so101_bimanual_franka_osc/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:99f93a7b5e3e4028c9dbc74789772346ef076d507971982e29ce52520e7510b8
+size 9269918

masters/embodiedswe/cutting_dice_franka_osc/meta.json

@@ -0,0 +1,52 @@
+{
+ "suite": "embodiedswe",
+ "task": "cutting_dice_franka_osc",
+ "task_dir": "embodiedswe-cutting-dice-franka-osc-i00-standard",
+ "source": "agent-success:codex-gpt6_astra-high-embodiedswe-cutting-dice-franka-osc-i00-standard/embodiedswe-cutting-dice-franka__QCU2Zao",
+ "agent": "Codex + GPT-6 Astra (high)",
+ "run": "codex-gpt6_astra-high",
+ "mode": "standard",
+ "trial": "QCU2Zao",
+ "trial_choice": "counted standard success; order: GPT-6 Astra > Claude Opus 5.5 > Gemini 4 Argon > Gemini 3.8 Flash, then the task's default mode (standard) before privileged",
+ "renderer": "isaac-rtx PathTracing spp=32 totalSpp=32 maxBounces=6 optix-denoiser firefly-filter (Isaac Sim 5.1 / Isaac Lab 2.3.2; dedicated render camera at the scene's own viewpoint, rendered natively at 2560x1440, focal 18.0 mm; recorded states posed, render only)",
+ "rendering": "path-tracing",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "frames": 775,
+ "seconds": 25.83,
+ "speed": 4.0,
+ "capture": "one recorded state every 2 control steps (0.133 s of simulated time) per video frame = 4x real time, 715 rendered frames; 0.5 s hold of the first frame, 1.5 s hold of the final frame",
+ "camera": {
+  "eye": [
+   0.72,
+   -0.62,
+   1.25
+  ],
+  "target": [
+   -0.08,
+   0.08,
+   0.95
+  ],
+  "focal_mm": 18.0,
+  "room_camera": [
+   [
+    0.72,
+    -0.62,
+    1.25
+   ],
+   [
+    -0.08,
+    0.08,
+    0.95
+   ]
+  ]
+ },
+ "success_check": true,
+ "success_check_how": "the video poses the graded episode's own recorded scene states (trajectory.npz frame_state: every asset's root pose and velocity and joint state, recorded during the graded live standard episode by the simulator service) on the task's own scene build (InteractiveScene.reset_to, no physics step), so it shows the graded run itself. Start state = the frozen instance (instance_ok 1, NVIDIA L40S). The task's success predicate on the posed final state: 0. Verdict (reward.json): success 1, progress 1.0.",
+ "image": "efe-private:embodiedswe-0.3.0 (sim service: efe:embodiedswe-standard-files-0.2.0)",
+ "core": true,
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, +faststart",
+ "poster": "the final frame",
+ "created": "2026-10-30"
+}

masters/embodiedswe/cutting_dice_franka_osc/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d1edd2dc418672b512155ca9acc78348efe48a028c4c897086c6638b1934aba8
+size 13708707

masters/embodiedswe/cutting_slice_franka_osc/meta.json

@@ -0,0 +1,52 @@
+{
+ "suite": "embodiedswe",
+ "task": "cutting_slice_franka_osc",
+ "task_dir": "embodiedswe-cutting-slice-franka-osc-i00-privileged",
+ "source": "agent-success:codex-gpt6_luna-xhigh-embodiedswe-cutting-slice-franka-osc-i00-privileged/embodiedswe-cutting-slice-franka__w9Df8br",
+ "agent": "Codex + GPT-6 Luna (xhigh)",
+ "run": "codex-gpt6_luna-xhigh",
+ "mode": "privileged",
+ "trial": "w9Df8br",
+ "trial_choice": "the only run that solved this task (GPT-6 Luna, counted privileged success); the other runs did not attempt it (screening: they run only tasks Luna fails in every mode)",
+ "renderer": "isaac-rtx PathTracing spp=32 totalSpp=32 maxBounces=6 optix-denoiser firefly-filter (Isaac Sim 5.1 / Isaac Lab 2.3.2; dedicated render camera at the scene's own viewpoint, rendered natively at 2560x1440, focal 18.0 mm; recorded states posed, render only)",
+ "rendering": "path-tracing",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "frames": 635,
+ "seconds": 21.17,
+ "speed": 8.0,
+ "capture": "one recorded state every 4 control steps (0.267 s of simulated time) per video frame = 8x real time, 575 rendered frames; 0.5 s hold of the first frame, 1.5 s hold of the final frame",
+ "camera": {
+  "eye": [
+   0.72,
+   -0.62,
+   1.25
+  ],
+  "target": [
+   -0.08,
+   0.08,
+   0.95
+  ],
+  "focal_mm": 18.0,
+  "room_camera": [
+   [
+    0.72,
+    -0.62,
+    1.25
+   ],
+   [
+    -0.08,
+    0.08,
+    0.95
+   ]
+  ]
+ },
+ "success_check": true,
+ "success_check_how": "the video poses the graded episode's own recorded scene states (trajectory.npz frame_state: every asset's root pose and velocity and joint state, recorded during the graded verifier rollout) on the task's own scene build (InteractiveScene.reset_to, no physics step), so it shows the graded run itself. Start state = the frozen instance (instance_ok 1, NVIDIA L40S). The task's success predicate on the posed final state: 0. Verdict (reward.json): success 1, progress 1.0. The verifier's own camera replay of the actions ended on the graded final state: camera_replay_match 1.",
+ "image": "efe-private:embodiedswe-0.3.0 (verifier tooling: efe:embodiedswe-files-0.3.1)",
+ "core": false,
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, +faststart",
+ "poster": "the final frame",
+ "created": "2026-10-30"
+}

masters/embodiedswe/cutting_slice_franka_osc/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f30cf9bb2511f6cb686078644656e9b9a0e3d956f269385e5e9137251f2abcbe
+size 13124712

masters/embodiedswe/highlights.json

@@ -0,0 +1,12 @@
+{"teaser": "puzzle_coffee_franka_osc",
+ "highlights": ["puzzle_coffee_franka_osc", "puzzle_syringe_bimanual_franka_osc", "assembly_pc_motherboard_franka_osc",
+                "puzzle_spatula_franka_osc", "packing_egg_carton_g1_pink_ik", "packing_tool_packing_franka_osc",
+                "puzzle_stack_blocks_g1_pink_ik"],
+ "why": {"puzzle_coffee_franka_osc": "a full capsule-coffee service in one program: open the bay, load the capsule, close it, stage the mug, start the brew, wait it out and serve the mug; GPT-6 Astra's counted success",
+         "puzzle_syringe_bimanual_franka_osc": "two arms draw liquid into a syringe and dose it; GPT-6 Astra's counted success",
+         "assembly_pc_motherboard_franka_osc": "seven bolts driven with an allen key inside a PC case, a long-horizon precision task",
+         "puzzle_spatula_franka_osc": "a tool-only rule: the bread is flipped and served with the spatula, never touched by the hand",
+         "packing_egg_carton_g1_pink_ik": "a humanoid places three eggs in a carton and closes the lid",
+         "packing_tool_packing_franka_osc": "three tools into three drawers: open, place, close, in sequence",
+         "puzzle_stack_blocks_g1_pink_ik": "a humanoid stacks blocks with its own hands"},
+ "why_fewer": ""}

masters/embodiedswe/locomanip_box_to_bin_g1_loco_pink_ik/meta.json

@@ -0,0 +1,52 @@
+{
+ "suite": "embodiedswe",
+ "task": "locomanip_box_to_bin_g1_loco_pink_ik",
+ "task_dir": "embodiedswe-locomanip-box-to-bin-g1-loco-pink-ik-i00-privileged",
+ "source": "agent-success:codex-gpt6_luna-xhigh-embodiedswe-locomanip-box-to-bin-g1-loco-pink-ik-i00-privileged/embodiedswe-locomanip-box-to-bin__7R4exsh",
+ "agent": "Codex + GPT-6 Luna (xhigh)",
+ "run": "codex-gpt6_luna-xhigh",
+ "mode": "privileged",
+ "trial": "7R4exsh",
+ "trial_choice": "the only run that solved this task (GPT-6 Luna, counted privileged success); the other runs did not attempt it (screening: they run only tasks Luna fails in every mode)",
+ "renderer": "isaac-rtx PathTracing spp=32 totalSpp=32 maxBounces=6 optix-denoiser firefly-filter (Isaac Sim 5.1 / Isaac Lab 2.3.2; dedicated render camera at the scene's own viewpoint, rendered natively at 2560x1440, focal 18.0 mm; recorded states posed, render only)",
+ "rendering": "path-tracing",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "frames": 755,
+ "seconds": 25.17,
+ "speed": 6.0,
+ "capture": "one recorded state every 10 control steps (0.200 s of simulated time) per video frame = 6x real time, 695 rendered frames; 0.5 s hold of the first frame, 1.5 s hold of the final frame",
+ "camera": {
+  "eye": [
+   2.9,
+   -2.7,
+   1.9
+  ],
+  "target": [
+   0.95,
+   0.2,
+   0.65
+  ],
+  "focal_mm": 18.0,
+  "room_camera": [
+   [
+    2.9,
+    -2.7,
+    1.9
+   ],
+   [
+    0.95,
+    0.2,
+    0.65
+   ]
+  ]
+ },
+ "success_check": true,
+ "success_check_how": "the video poses the graded episode's own recorded scene states (trajectory.npz frame_state: every asset's root pose and velocity and joint state, recorded during the graded verifier rollout) on the task's own scene build (InteractiveScene.reset_to, no physics step), so it shows the graded run itself. Start state = the frozen instance (instance_ok 1, NVIDIA L40S). The task's success predicate on the posed final state: 0. Verdict (reward.json): success 1, progress 1.0. The verifier's own camera replay of the actions ended on the graded final state: camera_replay_match 0.",
+ "image": "efe-private:embodiedswe-0.3.0 (verifier tooling: efe:embodiedswe-files-0.3.1)",
+ "core": false,
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, +faststart",
+ "poster": "the final frame",
+ "created": "2026-10-30"
+}

masters/embodiedswe/locomanip_box_to_bin_g1_loco_pink_ik/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:821324c4b48e037ab085580df28b4fd28a3ac72ec84c6d44aff5cd4e62d1104f
+size 6428106

masters/embodiedswe/locomanip_fruit_delivery_g1_loco_pink_ik/meta.json

@@ -0,0 +1,52 @@
+{
+ "suite": "embodiedswe",
+ "task": "locomanip_fruit_delivery_g1_loco_pink_ik",
+ "task_dir": "embodiedswe-locomanip-fruit-delivery-g1-loco-pink-ik-i00-privileged",
+ "source": "agent-success:codex-gpt6_luna-xhigh-embodiedswe-locomanip-fruit-delivery-g1-loco-pink-ik-i00-privileged/embodiedswe-locomanip-fruit-deli__svUzX8e",
+ "agent": "Codex + GPT-6 Luna (xhigh)",
+ "run": "codex-gpt6_luna-xhigh",
+ "mode": "privileged",
+ "trial": "svUzX8e",
+ "trial_choice": "the only run that solved this task (GPT-6 Luna, counted privileged success); the other runs did not attempt it (screening: they run only tasks Luna fails in every mode)",
+ "renderer": "isaac-rtx PathTracing spp=32 totalSpp=32 maxBounces=6 optix-denoiser firefly-filter (Isaac Sim 5.1 / Isaac Lab 2.3.2; dedicated render camera at the scene's own viewpoint, rendered natively at 2560x1440, focal 18.0 mm; recorded states posed, render only)",
+ "rendering": "path-tracing",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "frames": 869,
+ "seconds": 28.97,
+ "speed": 3.0,
+ "capture": "one recorded state every 6 control steps (0.100 s of simulated time) per video frame = 3x real time, 809 rendered frames; 0.5 s hold of the first frame, 1.5 s hold of the final frame",
+ "camera": {
+  "eye": [
+   3.5,
+   -4.5,
+   3.0
+  ],
+  "target": [
+   0.0,
+   0.0,
+   0.7
+  ],
+  "focal_mm": 18.0,
+  "room_camera": [
+   [
+    3.5,
+    -4.5,
+    3.0
+   ],
+   [
+    0.0,
+    0.0,
+    0.7
+   ]
+  ]
+ },
+ "success_check": true,
+ "success_check_how": "the video poses the graded episode's own recorded scene states (trajectory.npz frame_state: every asset's root pose and velocity and joint state, recorded during the graded verifier rollout) on the task's own scene build (InteractiveScene.reset_to, no physics step), so it shows the graded run itself. Start state = the frozen instance (instance_ok 1, NVIDIA L40S). The task's success predicate on the posed final state: 1. Verdict (reward.json): success 1, progress 1.0. The verifier's own camera replay of the actions ended on the graded final state: camera_replay_match 0.",
+ "image": "efe-private:embodiedswe-0.3.0 (verifier tooling: efe:embodiedswe-files-0.3.1)",
+ "core": false,
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, +faststart",
+ "poster": "the final frame",
+ "created": "2026-10-30"
+}

masters/embodiedswe/locomanip_fruit_delivery_g1_loco_pink_ik/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d9cd16e047f58ca6056265182ad60d04a23416cfd90ff9d3f1c20bc99444b821
+size 7692998

masters/embodiedswe/locomanip_wheel_carry_g1_loco_pink_ik/meta.json

@@ -0,0 +1,52 @@
+{
+ "suite": "embodiedswe",
+ "task": "locomanip_wheel_carry_g1_loco_pink_ik",
+ "task_dir": "embodiedswe-locomanip-wheel-carry-g1-loco-pink-ik-i00-privileged",
+ "source": "agent-success:codex-gpt6_luna-xhigh-embodiedswe-locomanip-wheel-carry-g1-loco-pink-ik-i00-privileged/embodiedswe-locomanip-wheel-carr__UcikH72",
+ "agent": "Codex + GPT-6 Luna (xhigh)",
+ "run": "codex-gpt6_luna-xhigh",
+ "mode": "privileged",
+ "trial": "UcikH72",
+ "trial_choice": "the only run that solved this task (GPT-6 Luna, counted privileged success); the other runs did not attempt it (screening: they run only tasks Luna fails in every mode)",
+ "renderer": "isaac-rtx PathTracing spp=32 totalSpp=32 maxBounces=6 optix-denoiser firefly-filter (Isaac Sim 5.1 / Isaac Lab 2.3.2; dedicated render camera at the scene's own viewpoint, rendered natively at 2560x1440, focal 18.0 mm; recorded states posed, render only)",
+ "rendering": "path-tracing",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "frames": 548,
+ "seconds": 18.27,
+ "speed": 3.0,
+ "capture": "one recorded state every 5 control steps (0.100 s of simulated time) per video frame = 3x real time, 488 rendered frames; 0.5 s hold of the first frame, 1.5 s hold of the final frame",
+ "camera": {
+  "eye": [
+   -1.5,
+   -2.5,
+   1.8
+  ],
+  "target": [
+   1.5,
+   0.3,
+   0.6
+  ],
+  "focal_mm": 18.0,
+  "room_camera": [
+   [
+    -1.5,
+    -2.5,
+    1.8
+   ],
+   [
+    1.5,
+    0.3,
+    0.6
+   ]
+  ]
+ },
+ "success_check": true,
+ "success_check_how": "the video poses the graded episode's own recorded scene states (trajectory.npz frame_state: every asset's root pose and velocity and joint state, recorded during the graded verifier rollout) on the task's own scene build (InteractiveScene.reset_to, no physics step), so it shows the graded run itself. Start state = the frozen instance (instance_ok 1, NVIDIA L40S). The task's success predicate on the posed final state: 0. Verdict (reward.json): success 1, progress 1.0. The verifier's own camera replay of the actions ended on the graded final state: camera_replay_match 0.",
+ "image": "efe-private:embodiedswe-0.3.0 (verifier tooling: efe:embodiedswe-files-0.3.1)",
+ "core": false,
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, +faststart",
+ "poster": "the final frame",
+ "created": "2026-10-30"
+}

masters/embodiedswe/locomanip_wheel_carry_g1_loco_pink_ik/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:159ba7896dd5abe7fd5d5d26ac6f457d3c009b51b74ce3fbed58a968131ea5a6
+size 7244374

masters/embodiedswe/packing_clear_organic_objects_g1_pink_ik/meta.json

@@ -0,0 +1,52 @@
+{
+ "suite": "embodiedswe",
+ "task": "packing_clear_organic_objects_g1_pink_ik",
+ "task_dir": "embodiedswe-packing-clear-organic-objects-g1-pink-ik-i00-privileged",
+ "source": "agent-success:codex-gpt6_luna-xhigh-embodiedswe-packing-clear-organic-objects-g1-pink-ik-i00-privileged/embodiedswe-packing-clear-organi__X4Gzjd7",
+ "agent": "Codex + GPT-6 Luna (xhigh)",
+ "run": "codex-gpt6_luna-xhigh",
+ "mode": "privileged",
+ "trial": "X4Gzjd7",
+ "trial_choice": "the only run that solved this task (GPT-6 Luna, counted privileged success); the other runs did not attempt it (screening: they run only tasks Luna fails in every mode)",
+ "renderer": "isaac-rtx PathTracing spp=32 totalSpp=32 maxBounces=6 optix-denoiser firefly-filter (Isaac Sim 5.1 / Isaac Lab 2.3.2; dedicated render camera at the scene's own viewpoint, rendered natively at 2560x1440, focal 18.0 mm; recorded states posed, render only)",
+ "rendering": "path-tracing",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "frames": 763,
+ "seconds": 25.43,
+ "speed": 3.0,
+ "capture": "one recorded state every 6 control steps (0.100 s of simulated time) per video frame = 3x real time, 703 rendered frames; 0.5 s hold of the first frame, 1.5 s hold of the final frame",
+ "camera": {
+  "eye": [
+   1.4,
+   -1.7,
+   1.78
+  ],
+  "target": [
+   0.0,
+   0.0,
+   0.88
+  ],
+  "focal_mm": 18.0,
+  "room_camera": [
+   [
+    1.4,
+    -1.7,
+    1.78
+   ],
+   [
+    0.0,
+    0.0,
+    0.88
+   ]
+  ]
+ },
+ "success_check": true,
+ "success_check_how": "the video poses the graded episode's own recorded scene states (trajectory.npz frame_state: every asset's root pose and velocity and joint state, recorded during the graded verifier rollout) on the task's own scene build (InteractiveScene.reset_to, no physics step), so it shows the graded run itself. Start state = the frozen instance (instance_ok 1, NVIDIA L40S). The task's success predicate on the posed final state: 1. Verdict (reward.json): success 1, progress 1.0. The verifier's own camera replay of the actions ended on the graded final state: camera_replay_match 0.",
+ "image": "efe-private:embodiedswe-0.3.0 (verifier tooling: efe:embodiedswe-files-0.3.1)",
+ "core": false,
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, +faststart",
+ "poster": "the final frame",
+ "created": "2026-10-30"
+}

masters/embodiedswe/packing_clear_organic_objects_g1_pink_ik/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:74062a89fa7f11d69c697022dcfb1ef043e3f6f312d1c3f0bdf47cee9e953a28
+size 12232748

masters/embodiedswe/packing_egg_carton_g1_pink_ik/meta.json

@@ -0,0 +1,52 @@
+{
+ "suite": "embodiedswe",
+ "task": "packing_egg_carton_g1_pink_ik",
+ "task_dir": "embodiedswe-packing-egg-carton-g1-pink-ik-i00-privileged",
+ "source": "agent-success:codex-gpt6_astra-high-embodiedswe-packing-egg-carton-g1-pink-ik-i00-privileged/embodiedswe-packing-egg-carton-g__65i4b65",
+ "agent": "Codex + GPT-6 Astra (high)",
+ "run": "codex-gpt6_astra-high",
+ "mode": "privileged",
+ "trial": "65i4b65",
+ "trial_choice": "counted privileged success; order: GPT-6 Astra > Claude Opus 5.5 > Gemini 4 Argon > Gemini 3.8 Flash, then the task's default mode (standard) before privileged",
+ "renderer": "isaac-rtx PathTracing spp=32 totalSpp=32 maxBounces=6 optix-denoiser firefly-filter (Isaac Sim 5.1 / Isaac Lab 2.3.2; dedicated render camera at the scene's own viewpoint, rendered natively at 2560x1440, focal 18.0 mm; recorded states posed, render only)",
+ "rendering": "path-tracing",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "frames": 772,
+ "seconds": 25.73,
+ "speed": 6.0,
+ "capture": "one recorded state every 12 control steps (0.200 s of simulated time) per video frame = 6x real time, 712 rendered frames; 0.5 s hold of the first frame, 1.5 s hold of the final frame",
+ "camera": {
+  "eye": [
+   1.25,
+   0.15,
+   1.35
+  ],
+  "target": [
+   0.0,
+   -0.42,
+   0.78
+  ],
+  "focal_mm": 18.0,
+  "room_camera": [
+   [
+    1.25,
+    0.15,
+    1.3499999999999999
+   ],
+   [
+    0.0,
+    -0.42,
+    0.78
+   ]
+  ]
+ },
+ "success_check": true,
+ "success_check_how": "the video poses the graded episode's own recorded scene states (trajectory.npz frame_state: every asset's root pose and velocity and joint state, recorded during the graded verifier rollout) on the task's own scene build (InteractiveScene.reset_to, no physics step), so it shows the graded run itself. Start state = the frozen instance (instance_ok 1, NVIDIA L40S). The task's success predicate on the posed final state: 1. Verdict (reward.json): success 1, progress 1.0. The verifier's own camera replay of the actions ended on the graded final state: camera_replay_match 0.",
+ "image": "efe-private:embodiedswe-0.3.0 (verifier tooling: efe:embodiedswe-files-0.3.1)",
+ "core": true,
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, +faststart",
+ "poster": "the final frame",
+ "created": "2026-10-30"
+}

masters/embodiedswe/packing_egg_carton_g1_pink_ik/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f0ae1f652b83f910b4db0b1a5133c33983e16c5a6c58c9a54615501c9baf5405
+size 12956346

masters/embodiedswe/packing_pen_holder_franka_osc/meta.json

@@ -0,0 +1,52 @@
+{
+ "suite": "embodiedswe",
+ "task": "packing_pen_holder_franka_osc",
+ "task_dir": "embodiedswe-packing-pen-holder-franka-osc-i00-standard",
+ "source": "agent-success:codex-gpt6_luna-xhigh-embodiedswe-packing-pen-holder-franka-osc-i00-standard/embodiedswe-packing-pen-holder-f__YhFgS3L",
+ "agent": "Codex + GPT-6 Luna (xhigh)",
+ "run": "codex-gpt6_luna-xhigh",
+ "mode": "standard",
+ "trial": "YhFgS3L",
+ "trial_choice": "the only run that solved this task (GPT-6 Luna, counted standard success); the other runs did not attempt it (screening: they run only tasks Luna fails in every mode)",
+ "renderer": "isaac-rtx PathTracing spp=32 totalSpp=32 maxBounces=6 optix-denoiser firefly-filter (Isaac Sim 5.1 / Isaac Lab 2.3.2; dedicated render camera at the scene's own viewpoint, rendered natively at 2560x1440, focal 18.0 mm; recorded states posed, render only)",
+ "rendering": "path-tracing",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "frames": 616,
+ "seconds": 20.53,
+ "speed": 4.0,
+ "capture": "one recorded state every 2 control steps (0.133 s of simulated time) per video frame = 4x real time, 556 rendered frames; 0.5 s hold of the first frame, 1.5 s hold of the final frame",
+ "camera": {
+  "eye": [
+   -1.2,
+   -1.05,
+   1.2
+  ],
+  "target": [
+   0.1,
+   0.05,
+   0.75
+  ],
+  "focal_mm": 18.0,
+  "room_camera": [
+   [
+    -1.2,
+    -1.05,
+    1.2
+   ],
+   [
+    0.1,
+    0.05,
+    0.75
+   ]
+  ]
+ },
+ "success_check": true,
+ "success_check_how": "the video poses the graded episode's own recorded scene states (trajectory.npz frame_state: every asset's root pose and velocity and joint state, recorded during the graded live standard episode by the simulator service) on the task's own scene build (InteractiveScene.reset_to, no physics step), so it shows the graded run itself. Start state = the frozen instance (instance_ok 1, NVIDIA L40S). The task's success predicate on the posed final state: 1. Verdict (reward.json): success 1, progress 1.0.",
+ "image": "efe-private:embodiedswe-0.3.0 (sim service: efe:embodiedswe-standard-files-0.2.0)",
+ "core": false,
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, +faststart",
+ "poster": "the final frame",
+ "created": "2026-10-30"
+}