Masters: roboolympics-match, mirrored after review

masters/roboolympics-match/badminton_duelrl_i01/meta.json

@@ -0,0 +1,47 @@
+{
+ "suite": "roboolympics-match",
+ "task": "badminton_duelrl_i01",
+ "title": "Badminton duel vs an RL opponent, two Franka arms",
+ "source": "agent-success:codex-gpt6_astra-high-roboolympics-match-badminton-duelrl-i01-standard-realtime/roboolympics-badminton-duelrl-i0__uhA6nrt",
+ "success_check": true,
+ "success_check_how": "the replay of the recorded run with the task's own simulator code reproduces the run's final state hash and score exactly, and the run's verdict is a win (a counted success, verifier reward 1)",
+ "agent": "GPT-6 Astra (Codex, high)",
+ "result": "Astra wins 11:4 against the task's RL opponent (seat B, a second Franka arm on its gantry)",
+ "trial_choice": "a counted success; the exam maintainer's pick for this task",
+ "renderer": "Blender 4.2.23 LTS Cycles path tracing (GPU OptiX), adaptive sampling (threshold 0.02, up to 128 spp), OpenImageDenoise; kinematic playback of the recorded simulation poses (no physics in Blender; the match's own geometry and colours, full-detail robot maker meshes on the same link frames); daylight: a sun lamp (2 degree disc) and a Nishita sky (no sun disc), a plain ground plane in the scene's ground colour; AgX view transform, Medium High Contrast look; motion blur 180 degrees; a tracking camera (fixed elevation, per-frame fit of the robot(s) and the ball, smoothed in time) with a slow eased 10-degree orbit, one continuous shot (the roboolympics-fun suite's look)",
+ "rendering": "path-tracing",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "frames": 416,
+ "seconds": 13.87,
+ "speed": 1.0,
+ "clip": "GPT-6 Astra (near arm) wins a 12 s rally against the task's RL opponent (far arm) on the way to 11:4",
+ "camera": {
+  "mode": "fun-track",
+  "az_deg": -160,
+  "el_deg": 22.0,
+  "fovy_deg": 34,
+  "fill": 0.72,
+  "orbit_deg": 10,
+  "distance_m": [
+   9.46,
+   10.683
+  ]
+ },
+ "revision": "badminton-wallclock-12+duel-14+core-0.2",
+ "holds": {
+  "start_s": 0.5,
+  "end_s": 1.5,
+  "note": "first/last frame held, no extra simulation"
+ },
+ "visual_changes": [
+  "daylight sky and a plain ground plane in the scene's ground colour (drawn-only)",
+  "the shuttle drawn as a real feather shuttlecock at the match shuttle's position, cork forward along its recorded velocity (the simulation uses a point-mass shuttle)",
+  "the racket drawn as a real badminton racket on the match racket's frame"
+ ],
+ "poster": "plan frame 333: the key moment of the clip",
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, 30 fps constant, +faststart, no audio, no text or overlays",
+ "created": "2026-10-19",
+ "contributor": "EAO team (Chi Gui), Embodied Agent Olympics"
+}

masters/roboolympics-match/badminton_duelrl_i01/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6dc037226cf5d6fa7656334e397ee1e28fbbf7ba7926f25ba92b79cab27c5aa4
+size 7094485

masters/roboolympics-match/badminton_match_i01/meta.json

@@ -0,0 +1,51 @@
+{
+ "suite": "roboolympics-match",
+ "task": "badminton_match_i01",
+ "title": "Badminton match, Franka arm (instance 1)",
+ "source": "agent-success:codex-gpt6_astra-high-roboolympics-match-badminton-match-i01-standard-realtime/roboolympics-badminton-match-i01__jPV4NA8",
+ "success_check": true,
+ "success_check_how": "the replay of the recorded run with the task's own simulator code reproduces the run's final state hash and score exactly, and the run's verdict is a win (a counted success, verifier reward 1)",
+ "agent": "GPT-6 Astra (Codex, high)",
+ "result": "Astra wins 11:4 against the built-in level-1 opponent",
+ "trial_choice": "a counted success; the exam maintainer's pick for this task",
+ "renderer": "Blender 4.2.23 LTS Cycles path tracing (GPU OptiX), adaptive sampling (threshold 0.02, up to 128 spp), OpenImageDenoise; kinematic playback of the recorded simulation poses (no physics in Blender; the match's own geometry and colours, full-detail robot maker meshes on the same link frames); daylight: a sun lamp (2 degree disc) and a Nishita sky (no sun disc), a plain ground plane in the scene's ground colour; AgX view transform, Medium High Contrast look; motion blur 180 degrees; a tracking camera (fixed elevation, per-frame fit of the robot(s) and the ball, smoothed in time) with a slow eased 10-degree orbit, one continuous shot (the roboolympics-fun suite's look)",
+ "rendering": "path-tracing",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "frames": 436,
+ "seconds": 14.53,
+ "speed": 1.0,
+ "clip": "GPT-6 Astra wins a 28 s rally against the built-in level-1 opponent (3:1); the clip shows its last 12 s, ending with Astra's winner landing in the opponent's court; the opponent arm is drawn for display only",
+ "camera": {
+  "mode": "fun-track",
+  "az_deg": -160,
+  "el_deg": 22.0,
+  "fovy_deg": 34,
+  "fill": 0.85,
+  "orbit_deg": 10,
+  "distance_m": [
+   7.742,
+   9.794
+  ]
+ },
+ "revision": "badminton-wallclock-12+duel-14+core-0.2",
+ "holds": {
+  "start_s": 0.5,
+  "end_s": 1.5,
+  "note": "first/last frame held, no extra simulation"
+ },
+ "visual_changes": [
+  "daylight sky and a plain ground plane in the scene's ground colour (drawn-only)",
+  "the shuttle drawn as a real feather shuttlecock at the match shuttle's position, cork forward along its recorded velocity (the simulation uses a point-mass shuttle)",
+  "the racket drawn as a real badminton racket on the match racket's frame",
+  "an opponent arm (the same arm and racket as the player's, mirrored on the far side) drawn for display only: posed by kinematics (inverse kinematics to each recorded return of the scripted opponent, swing and ready poses between them); the ball path is the recording's; in the task the scripted opponent has no body"
+ ],
+ "notes": [
+  "opponent arm drawn for display only; in the task the scripted opponent has no body"
+ ],
+ "poster": "plan frame 351: the key moment of the clip",
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, 30 fps constant, +faststart, no audio, no text or overlays",
+ "created": "2026-10-19",
+ "contributor": "EAO team (Chi Gui), Embodied Agent Olympics"
+}

masters/roboolympics-match/badminton_match_i01/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2f16a31f337c4866f2c6f1d4e5f0a2b25a84ec53144b549eb9d33e2d249770d7
+size 8046331

masters/roboolympics-match/badminton_match_i02/meta.json

@@ -0,0 +1,51 @@
+{
+ "suite": "roboolympics-match",
+ "task": "badminton_match_i02",
+ "title": "Badminton match, Franka arm (instance 2)",
+ "source": "agent-success:codex-gpt6_astra-high-roboolympics-match-badminton-match-i02-standard-realtime/roboolympics-badminton-match-i02__tDdNAfB",
+ "success_check": true,
+ "success_check_how": "the replay of the recorded run with the task's own simulator code reproduces the run's final state hash and score exactly, and the run's verdict is a win (a counted success, verifier reward 1)",
+ "agent": "GPT-6 Astra (Codex, high)",
+ "result": "Astra wins 11:8 against the built-in level-2 opponent",
+ "trial_choice": "a counted success; the exam maintainer's pick for this task",
+ "renderer": "Blender 4.2.23 LTS Cycles path tracing (GPU OptiX), adaptive sampling (threshold 0.02, up to 128 spp), OpenImageDenoise; kinematic playback of the recorded simulation poses (no physics in Blender; the match's own geometry and colours, full-detail robot maker meshes on the same link frames); daylight: a sun lamp (2 degree disc) and a Nishita sky (no sun disc), a plain ground plane in the scene's ground colour; AgX view transform, Medium High Contrast look; motion blur 180 degrees; a tracking camera (fixed elevation, per-frame fit of the robot(s) and the ball, smoothed in time) with a slow eased 10-degree orbit, one continuous shot (the roboolympics-fun suite's look)",
+ "rendering": "path-tracing",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "frames": 602,
+ "seconds": 20.07,
+ "speed": 1.0,
+ "clip": "the last two points of GPT-6 Astra's 11:8 win over the built-in level-2 opponent (10:8, then match point 11:8); the opponent arm is drawn for display only",
+ "camera": {
+  "mode": "fun-track",
+  "az_deg": -160,
+  "el_deg": 22.0,
+  "fovy_deg": 34,
+  "fill": 0.85,
+  "orbit_deg": 10,
+  "distance_m": [
+   7.727,
+   10.022
+  ]
+ },
+ "revision": "badminton-wallclock-12+duel-14+core-0.2",
+ "holds": {
+  "start_s": 0.5,
+  "end_s": 1.5,
+  "note": "first/last frame held, no extra simulation"
+ },
+ "visual_changes": [
+  "daylight sky and a plain ground plane in the scene's ground colour (drawn-only)",
+  "the shuttle drawn as a real feather shuttlecock at the match shuttle's position, cork forward along its recorded velocity (the simulation uses a point-mass shuttle)",
+  "the racket drawn as a real badminton racket on the match racket's frame",
+  "an opponent arm (the same arm and racket as the player's, mirrored on the far side) drawn for display only: posed by kinematics (inverse kinematics to each recorded return of the scripted opponent, swing and ready poses between them); the ball path is the recording's; in the task the scripted opponent has no body"
+ ],
+ "notes": [
+  "opponent arm drawn for display only; in the task the scripted opponent has no body"
+ ],
+ "poster": "plan frame 525: the key moment of the clip",
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, 30 fps constant, +faststart, no audio, no text or overlays",
+ "created": "2026-10-19",
+ "contributor": "EAO team (Chi Gui), Embodied Agent Olympics"
+}

masters/roboolympics-match/badminton_match_i02/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1974a18eb419223c8b5cfbade07cb17121e5b4e45e6e659952b83f4d1f634483
+size 11670850

masters/roboolympics-match/badminton_match_i04/meta.json

@@ -0,0 +1,51 @@
+{
+ "suite": "roboolympics-match",
+ "task": "badminton_match_i04",
+ "title": "Badminton match, Franka arm (instance 4)",
+ "source": "scene-orbit",
+ "shows": "oracle-attempt (lost): the task's reference solution: it loses 3:11 to the built-in level-5 opponent",
+ "why_no_video": "No evaluated agent has won this task (best: GPT-6 Astra 8:11), and the task's reference solution does not win it either. As for the roboolympics-fun suite's unsolved tasks, this video is that reference solution's match replayed exactly (not a solved run), showing its best moment: the reference solution loses 3:11 to the built-in level-5 opponent.",
+ "success_check": false,
+ "success_check_how": "the match's own result at the end of the replay (a loss), equal to the recording",
+ "oracle_recording": "played by the task's reference player with the task's own simulator code and configuration (seed, level, action level); recorded in-process (no network, no wall clock), the same player the task's oracle runs over its HTTP interface",
+ "renderer": "Blender 4.2.23 LTS Cycles path tracing (GPU OptiX), adaptive sampling (threshold 0.02, up to 128 spp), OpenImageDenoise; kinematic playback of the recorded simulation poses (no physics in Blender; the match's own geometry and colours, full-detail robot maker meshes on the same link frames); daylight: a sun lamp (2 degree disc) and a Nishita sky (no sun disc), a plain ground plane in the scene's ground colour; AgX view transform, Medium High Contrast look; motion blur 180 degrees; a tracking camera (fixed elevation, per-frame fit of the robot(s) and the ball, smoothed in time) with a slow eased 10-degree orbit, one continuous shot (the roboolympics-fun suite's look)",
+ "rendering": "path-tracing",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "frames": 593,
+ "seconds": 19.77,
+ "speed": 1.0,
+ "clip": "the task's reference solution wins its longest rally (17.6 s, 2:4) against the built-in level-5 opponent; it loses the match 3:11; the opponent arm is drawn for display only",
+ "camera": {
+  "mode": "fun-track",
+  "az_deg": -160,
+  "el_deg": 22.0,
+  "fovy_deg": 34,
+  "fill": 0.85,
+  "orbit_deg": 10,
+  "distance_m": [
+   7.904,
+   9.654
+  ]
+ },
+ "revision": "badminton-wallclock-12+duel-14+core-0.2",
+ "holds": {
+  "start_s": 0.5,
+  "end_s": 1.5,
+  "note": "first/last frame held, no extra simulation"
+ },
+ "visual_changes": [
+  "daylight sky and a plain ground plane in the scene's ground colour (drawn-only)",
+  "the shuttle drawn as a real feather shuttlecock at the match shuttle's position, cork forward along its recorded velocity (the simulation uses a point-mass shuttle)",
+  "the racket drawn as a real badminton racket on the match racket's frame",
+  "an opponent arm (the same arm and racket as the player's, mirrored on the far side) drawn for display only: posed by kinematics (inverse kinematics to each recorded return of the scripted opponent, swing and ready poses between them); the ball path is the recording's; in the task the scripted opponent has no body"
+ ],
+ "notes": [
+  "opponent arm drawn for display only; in the task the scripted opponent has no body"
+ ],
+ "poster": "plan frame 513: the key moment of the clip",
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, 30 fps constant, +faststart, no audio, no text or overlays",
+ "created": "2026-10-19",
+ "contributor": "EAO team (Chi Gui), Embodied Agent Olympics"
+}

masters/roboolympics-match/badminton_match_i04/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:631e701ae4637e51c7d30e81952ccef82d184e99d6573ab87f6d5b29c1641841
+size 11456237

masters/roboolympics-match/badminton_matchlow_i01/meta.json

@@ -0,0 +1,51 @@
+{
+ "suite": "roboolympics-match",
+ "task": "badminton_matchlow_i01",
+ "title": "Badminton match, low-level control (instance 1)",
+ "source": "scene-orbit",
+ "shows": "oracle-attempt (lost): the task's reference solution: it (its act skill in the loop) loses 8:11 to the built-in level-2 opponent",
+ "why_no_video": "No evaluated agent has won this task (best: GPT-6 Astra 2:11), and the task's reference solution does not win it either. As for the roboolympics-fun suite's unsolved tasks, this video is that reference solution's match replayed exactly (not a solved run), showing its best moment: the reference solution (its act skill in the loop) loses 8:11 to the built-in level-2 opponent.",
+ "success_check": false,
+ "success_check_how": "the match's own result at the end of the replay (a loss), equal to the recording",
+ "oracle_recording": "played by the task's reference player with the task's own simulator code and configuration (seed, level, action level); recorded in-process (no network, no wall clock), the same player the task's oracle runs over its HTTP interface",
+ "renderer": "Blender 4.2.23 LTS Cycles path tracing (GPU OptiX), adaptive sampling (threshold 0.02, up to 128 spp), OpenImageDenoise; kinematic playback of the recorded simulation poses (no physics in Blender; the match's own geometry and colours, full-detail robot maker meshes on the same link frames); daylight: a sun lamp (2 degree disc) and a Nishita sky (no sun disc), a plain ground plane in the scene's ground colour; AgX view transform, Medium High Contrast look; motion blur 180 degrees; a tracking camera (fixed elevation, per-frame fit of the robot(s) and the ball, smoothed in time) with a slow eased 10-degree orbit, one continuous shot (the roboolympics-fun suite's look)",
+ "rendering": "path-tracing",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "frames": 584,
+ "seconds": 19.47,
+ "speed": 1.0,
+ "clip": "the task's low-level reference solution (its own act skill in the loop) wins a 22 s rally (3:2) against the built-in level-2 opponent; it loses the match 8:11; the opponent arm is drawn for display only",
+ "camera": {
+  "mode": "fun-track",
+  "az_deg": -160,
+  "el_deg": 22.0,
+  "fovy_deg": 34,
+  "fill": 0.85,
+  "orbit_deg": 10,
+  "distance_m": [
+   8.492,
+   10.669
+  ]
+ },
+ "revision": "badminton-wallclock-12+duel-14+core-0.2",
+ "holds": {
+  "start_s": 0.5,
+  "end_s": 1.5,
+  "note": "first/last frame held, no extra simulation"
+ },
+ "visual_changes": [
+  "daylight sky and a plain ground plane in the scene's ground colour (drawn-only)",
+  "the shuttle drawn as a real feather shuttlecock at the match shuttle's position, cork forward along its recorded velocity (the simulation uses a point-mass shuttle)",
+  "the racket drawn as a real badminton racket on the match racket's frame",
+  "an opponent arm (the same arm and racket as the player's, mirrored on the far side) drawn for display only: posed by kinematics (inverse kinematics to each recorded return of the scripted opponent, swing and ready poses between them); the ball path is the recording's; in the task the scripted opponent has no body"
+ ],
+ "notes": [
+  "opponent arm drawn for display only; in the task the scripted opponent has no body"
+ ],
+ "poster": "plan frame 507: the key moment of the clip",
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, 30 fps constant, +faststart, no audio, no text or overlays",
+ "created": "2026-10-19",
+ "contributor": "EAO team (Chi Gui), Embodied Agent Olympics"
+}

masters/roboolympics-match/badminton_matchlow_i01/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ce6c9ccc4b687d7be5a08a3ccb7b5ed5fe95850c834a2743cce79ad4274db05f
+size 10704023

masters/roboolympics-match/fencing_bout_i01/meta.json

@@ -0,0 +1,48 @@
+{
+ "suite": "roboolympics-match",
+ "task": "fencing_bout_i01",
+ "title": "Fencing bout, two Unitree G1 (instance 1)",
+ "source": "agent-success:claude_code-opus_5_5-high-roboolympics-match-fencing-bout-i01-standard/roboolympics-fencing-bout-i01-st__JxrzR2i",
+ "success_check": true,
+ "success_check_how": "the replay of the recorded run with the task's own simulator code reproduces the run's final state hash and score exactly, and the run's verdict is a win (a counted success, verifier reward 1)",
+ "agent": "Claude Opus 5.5 (Claude Code, high)",
+ "result": "Opus wins 15:14 against the built-in level-1 fencer (28 phrases)",
+ "trial_choice": "a counted success; the exam maintainer's pick for this task",
+ "renderer": "Blender 4.2.23 LTS Cycles path tracing (GPU OptiX), adaptive sampling (threshold 0.02, up to 128 spp), OpenImageDenoise; kinematic playback of the recorded simulation poses (no physics in Blender; the match's own geometry and colours, full-detail robot maker meshes on the same link frames); daylight: a sun lamp (2 degree disc) and a Nishita sky (no sun disc), a plain ground plane in the scene's ground colour; AgX view transform, Medium High Contrast look; motion blur 180 degrees; a tracking camera (fixed elevation, per-frame fit of the robot(s) and the ball, smoothed in time) with a slow eased 10-degree orbit, one continuous shot (the roboolympics-fun suite's look)",
+ "rendering": "path-tracing",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "frames": 325,
+ "seconds": 10.83,
+ "speed": 1.0,
+ "speed_note": "phrase motion at match speed (10 ms steps); the on-guard / halt holds inside the timeline are display holds (no simulation), as in our approved fencing_bout clip",
+ "clip": "Claude Opus 5.5 (seat A, blue) wins the bout 15:14 against the level-1 scripted fencer: the last three phrases, three touches in a row by A (12:14 -> 13:14 -> 14:14 -> 15:14), the last one the winning touch; each phrase is the sport's looked-up phrase re-played per 10 ms on the whole-body controller the phrase table was generated with (exact at the stored poses), shown with an on-guard hold before (0.8 s) and a hold of the halt after (1.0 s); between phrases the sport resets both fencers to on guard (a cut inside the one shot)",
+ "camera": {
+  "mode": "fun-track",
+  "az_deg": -80,
+  "el_deg": 22.0,
+  "fovy_deg": 34,
+  "fill": 0.9,
+  "orbit_deg": 10,
+  "distance_m": [
+   3.617,
+   4.118
+  ]
+ },
+ "revision": "fencing-7+core-0.2 (exam phrases/table.json + frames.npz byte-identical to main 24aa3286, the table FENCEREND re-plays)",
+ "holds": {
+  "start_s": 0.5,
+  "end_s": 1.5,
+  "note": "first/last frame held, no extra simulation"
+ },
+ "visual_changes": [
+  "daylight sky and a plain ground plane in the scene's ground colour (drawn-only)",
+  "the sword hand is not drawn (in the simulation it is display-only geometry, no contact); a wrist flange holds the sword",
+  "the motion between the three poses per phrase the match stores is the task's own whole-body controller re-played at 10 ms (exact against the stored poses); the fencers are put back on guard between phrases (a cut in the shot)"
+ ],
+ "poster": "plan frame 234: the key moment of the clip",
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, 30 fps constant, +faststart, no audio, no text or overlays",
+ "created": "2026-10-19",
+ "contributor": "EAO team (Chi Gui), Embodied Agent Olympics"
+}

masters/roboolympics-match/fencing_bout_i01/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d74d03b8b9f815f2972aad2be496d06a6bb2d865089ffb008871718b43e65eb4
+size 6040112

masters/roboolympics-match/fencing_bout_i02/meta.json

@@ -0,0 +1,48 @@
+{
+ "suite": "roboolympics-match",
+ "task": "fencing_bout_i02",
+ "title": "Fencing bout, two Unitree G1 (instance 2)",
+ "source": "agent-success:codex-gpt6_luna-xhigh-roboolympics-match-fencing-bout-i02-standard/roboolympics-fencing-bout-i02-st__czQaDQo",
+ "success_check": true,
+ "success_check_how": "the replay of the recorded run with the task's own simulator code reproduces the run's final state hash and score exactly, and the run's verdict is a win (a counted success, verifier reward 1)",
+ "agent": "GPT-6 Luna (Codex, xhigh)",
+ "result": "Luna wins 15:13 against the built-in level-2 fencer (30 phrases)",
+ "trial_choice": "a counted success; the exam maintainer's pick for this task",
+ "renderer": "Blender 4.2.23 LTS Cycles path tracing (GPU OptiX), adaptive sampling (threshold 0.02, up to 128 spp), OpenImageDenoise; kinematic playback of the recorded simulation poses (no physics in Blender; the match's own geometry and colours, full-detail robot maker meshes on the same link frames); daylight: a sun lamp (2 degree disc) and a Nishita sky (no sun disc), a plain ground plane in the scene's ground colour; AgX view transform, Medium High Contrast look; motion blur 180 degrees; a tracking camera (fixed elevation, per-frame fit of the robot(s) and the ball, smoothed in time) with a slow eased 10-degree orbit, one continuous shot (the roboolympics-fun suite's look)",
+ "rendering": "path-tracing",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "frames": 310,
+ "seconds": 10.33,
+ "speed": 1.0,
+ "speed_note": "phrase motion at match speed (10 ms steps); the on-guard / halt holds inside the timeline are display holds (no simulation), as in our approved fencing_bout clip",
+ "clip": "GPT-6 Luna (seat A, blue) wins the bout 15:13 against the level-2 scripted fencer: the last three phrases, touch A (13:12 -> 14:12), touch B (14:13), then A's winning touch (15:13); each phrase re-played per 10 ms on the whole-body controller the phrase table was generated with (exact at the stored poses), drawn as the sport's own images draw it (half-turned so A stays on its own end), with an on-guard hold before (0.8 s) and a hold of the halt after (1.0 s); between phrases the sport resets both fencers to on guard (a cut inside the one shot)",
+ "camera": {
+  "mode": "fun-track",
+  "az_deg": -80,
+  "el_deg": 22.0,
+  "fovy_deg": 34,
+  "fill": 0.9,
+  "orbit_deg": 10,
+  "distance_m": [
+   3.424,
+   4.15
+  ]
+ },
+ "revision": "fencing-7+core-0.2 (exam phrases/table.json + frames.npz byte-identical to main 24aa3286, the table FENCEREND re-plays)",
+ "holds": {
+  "start_s": 0.5,
+  "end_s": 1.5,
+  "note": "first/last frame held, no extra simulation"
+ },
+ "visual_changes": [
+  "daylight sky and a plain ground plane in the scene's ground colour (drawn-only)",
+  "the sword hand is not drawn (in the simulation it is display-only geometry, no contact); a wrist flange holds the sword",
+  "the motion between the three poses per phrase the match stores is the task's own whole-body controller re-played at 10 ms (exact against the stored poses); the fencers are put back on guard between phrases (a cut in the shot)"
+ ],
+ "poster": "plan frame 219: the key moment of the clip",
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, 30 fps constant, +faststart, no audio, no text or overlays",
+ "created": "2026-10-19",
+ "contributor": "EAO team (Chi Gui), Embodied Agent Olympics"
+}

masters/roboolympics-match/fencing_bout_i02/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3002088b2a7cdc7547b54509f4d79f6acb2069f109d441ae09f87b29fe865290
+size 5786328

masters/roboolympics-match/foosball_match_i01/meta.json

@@ -0,0 +1,46 @@
+{
+ "suite": "roboolympics-match",
+ "task": "foosball_match_i01",
+ "title": "Foosball match (instance 1)",
+ "source": "agent-success:codex-gpt6_astra-high-roboolympics-match-foosball-match-i01-standard-realtime/roboolympics-foosball-match-i01__fieEuq8",
+ "success_check": true,
+ "success_check_how": "the replay of the recorded run with the task's own simulator code reproduces the run's final state hash and score exactly, and the run's verdict is a win (a counted success, verifier reward 1)",
+ "agent": "GPT-6 Astra (Codex, high)",
+ "result": "Astra wins 5:3 against the built-in level-3 opponent (low-level actions)",
+ "trial_choice": "a counted success; the exam maintainer's pick for this task",
+ "renderer": "Blender 4.2.23 LTS Cycles path tracing (GPU OptiX), adaptive sampling (threshold 0.02, up to 128 spp), OpenImageDenoise; kinematic playback of the recorded simulation poses (no physics in Blender; the match's own geometry and colours, full-detail robot maker meshes on the same link frames); daylight: a sun lamp (2 degree disc) and a Nishita sky (no sun disc), a plain ground plane in the scene's ground colour; AgX view transform, Medium High Contrast look; motion blur 180 degrees; a tracking camera (fixed elevation, per-frame fit of the robot(s) and the ball, smoothed in time) with a slow eased 10-degree orbit, one continuous shot (the roboolympics-fun suite's look)",
+ "rendering": "path-tracing",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "frames": 373,
+ "seconds": 12.43,
+ "speed": 1.0,
+ "clip": "GPT-6 Astra's program (blue rods) against the built-in level-3 opponent: the last 10.4 s of the match, ending with Astra's winning goal for 5:3",
+ "camera": {
+  "mode": "fun-track",
+  "az_deg": -90,
+  "el_deg": 40.0,
+  "fovy_deg": 34,
+  "fill": 0.78,
+  "orbit_deg": 10,
+  "distance_m": [
+   1.959,
+   1.959
+  ]
+ },
+ "revision": "foosball-program-v2+core-0.2",
+ "holds": {
+  "start_s": 0.5,
+  "end_s": 1.5,
+  "note": "first/last frame held, no extra simulation"
+ },
+ "visual_changes": [
+  "daylight sky and a plain ground plane in the scene's ground colour (drawn-only)",
+  "the table cabinet and legs drawn under the playfield (display only)"
+ ],
+ "poster": "plan frame 312: the key moment of the clip",
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, 30 fps constant, +faststart, no audio, no text or overlays",
+ "created": "2026-10-19",
+ "contributor": "EAO team (Chi Gui), Embodied Agent Olympics"
+}

masters/roboolympics-match/foosball_match_i01/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:289a584f187205d50f05be01e2d032cf9e3d3805ba41b817beff201d26994d95
+size 8956609

masters/roboolympics-match/highlights.json

@@ -0,0 +1,36 @@
+{
+ "teaser": "soccer_jointgpu_i01",
+ "highlights": [
+  "soccer_jointgpu_i01",
+  "fencing_bout_i01",
+  "soccer_match",
+  "badminton_match_i01",
+  "tabletennis_match_i01",
+  "foosball_match_i01",
+  "badminton_duelrl_i01",
+  "fencing_bout_i02"
+ ],
+ "why_fewer": "",
+ "why": {
+  "soccer_jointgpu_i01": "two Booster T1 humanoids under joint-level control beat a level-5 RL team 13:8 (GPT-6 Astra)",
+  "fencing_bout_i01": "a Unitree G1 fencing bout won 15:14 on the last touch (Claude Opus 5.5)",
+  "soccer_match": "humanoid 2 v 2 soccer, a 9:0 win over the level-2 team (GPT-6 Astra)",
+  "badminton_match_i01": "a Franka arm on a gantry wins a 28 s rally and the match 11:4 (GPT-6 Astra)",
+  "tabletennis_match_i01": "a Franka arm wins the longest rally and the match 6:4 against the level-2 opponent (GPT-6 Astra)",
+  "foosball_match_i01": "foosball with low-level rod control, the winning goal of a 5:3 (GPT-6 Astra)",
+  "badminton_duelrl_i01": "two Franka arms: the agent beats the task's RL opponent 11:4 (GPT-6 Astra)",
+  "fencing_bout_i02": "a Unitree G1 fencing bout won 15:13 (GPT-6 Luna)"
+ },
+ "teaser_cuts": {
+  "tasks": [
+   "soccer_jointgpu_i01",
+   "fencing_bout_i01",
+   "badminton_match_i01",
+   "tabletennis_match_i01",
+   "foosball_match_i01",
+   "soccer_match"
+  ],
+  "seconds_each": 4.5,
+  "note": "the last 4.5 s before the end hold of each video, hard cuts"
+ }
+}

masters/roboolympics-match/soccer_jointcpu_i01/meta.json

@@ -0,0 +1,47 @@
+{
+ "suite": "roboolympics-match",
+ "task": "soccer_jointcpu_i01",
+ "title": "Humanoid soccer 2 v 2, joint-level control (CPU tier)",
+ "source": "scene-orbit",
+ "shows": "oracle-attempt (lost): the task's reference solution: it loses 0:10 to the built-in level-5 team (mercy rule at 366.8 s)",
+ "why_no_video": "No evaluated agent has won this task (best: every evaluated agent lost 0:10 or scored none), and the task's reference solution does not win it either. As for the roboolympics-fun suite's unsolved tasks, this video is that reference solution's match replayed exactly (not a solved run), showing its best moment: the reference solution loses 0:10 to the built-in level-5 team (mercy rule at 366.8 s).",
+ "success_check": false,
+ "success_check_how": "the match's own result at the end of the replay (a loss), equal to the recording",
+ "oracle_recording": "played by the task's reference player with the task's own simulator code and configuration (seed, level, action level); an oracle run of the task in its evaluation harness",
+ "renderer": "Blender 4.2.23 LTS Cycles path tracing (GPU OptiX), adaptive sampling (threshold 0.02, up to 128 spp), OpenImageDenoise; kinematic playback of the recorded simulation poses (no physics in Blender; the match's own geometry and colours, full-detail robot maker meshes on the same link frames); daylight: a sun lamp (2 degree disc) and a Nishita sky (no sun disc), a plain ground plane in the scene's ground colour; AgX view transform, Medium High Contrast look; motion blur 180 degrees; a tracking camera (fixed elevation, per-frame fit of the robot(s) and the ball, smoothed in time) with a slow eased 10-degree orbit, one continuous shot (the roboolympics-fun suite's look)",
+ "rendering": "path-tracing",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "frames": 379,
+ "seconds": 12.63,
+ "speed": 1.0,
+ "clip": "the task's oracle (reference joint-level team, blue) vs the built-in level-5 RL team (red), match time 26.9-37.5 s: the oracle's robots get back up and contest the ball near the centre; the oracle lost 0:10, this is its best stretch (no oracle shot in the match)",
+ "camera": {
+  "mode": "fun-track",
+  "az_deg": -100,
+  "el_deg": 22.0,
+  "fovy_deg": 34,
+  "fill": 0.8,
+  "orbit_deg": 10,
+  "distance_m": [
+   8.462,
+   13.151
+  ]
+ },
+ "revision": "soccer-2v2-11+core-0.2 (source_sha header 95d52171)",
+ "holds": {
+  "start_s": 0.5,
+  "end_s": 1.5,
+  "note": "first/last frame held, no extra simulation"
+ },
+ "visual_changes": [
+  "daylight sky and a plain ground plane in the scene's ground colour (drawn-only)",
+  "the ball drawn white (orange in the match)",
+  "display cord nets in place of the match's box goal nets"
+ ],
+ "poster": "plan frame 54: the key moment of the clip",
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, 30 fps constant, +faststart, no audio, no text or overlays",
+ "created": "2026-10-19",
+ "contributor": "EAO team (Chi Gui), Embodied Agent Olympics"
+}

masters/roboolympics-match/soccer_jointcpu_i01/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:adcee7ed6b029f29255884f63224adabb9069c896d798301effb0cea8ad87958
+size 8680322

masters/roboolympics-match/soccer_jointgpu_i01/meta.json

@@ -0,0 +1,47 @@
+{
+ "suite": "roboolympics-match",
+ "task": "soccer_jointgpu_i01",
+ "title": "Humanoid soccer 2 v 2, joint-level control (GPU tier)",
+ "source": "agent-success:codex-gpt6_astra-high-roboolympics-match-soccer-jointgpu-i01-standard-realtime/roboolympics-soccer-jointgpu-i01-standard-realtime__ZJZf2Gs",
+ "success_check": true,
+ "success_check_how": "the replay of the recorded run with the task's own simulator code reproduces the run's final state hash and score exactly, and the run's verdict is a win (a counted success, verifier reward 1)",
+ "agent": "GPT-6 Astra (Codex, high)",
+ "result": "Astra wins 13:8 against the built-in level-5 team (RL joint-level team)",
+ "trial_choice": "a counted success; the exam maintainer's pick for this task",
+ "renderer": "Blender 4.2.23 LTS Cycles path tracing (GPU OptiX), adaptive sampling (threshold 0.02, up to 128 spp), OpenImageDenoise; kinematic playback of the recorded simulation poses (no physics in Blender; the match's own geometry and colours, full-detail robot maker meshes on the same link frames); daylight: a sun lamp (2 degree disc) and a Nishita sky (no sun disc), a plain ground plane in the scene's ground colour; AgX view transform, Medium High Contrast look; motion blur 180 degrees; a tracking camera (fixed elevation, per-frame fit of the robot(s) and the ball, smoothed in time) with a slow eased 10-degree orbit, one continuous shot (the roboolympics-fun suite's look)",
+ "rendering": "path-tracing",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "frames": 520,
+ "seconds": 17.33,
+ "speed": 1.0,
+ "clip": "GPT-6 Astra's joint-level team (blue) vs the built-in level-5 RL team (red), match time 218.6-233.9 s: from the kick-off after the 4:2 goal, A1 wins the ball and scores (5:2 at 233.92 s); Astra won 13:8",
+ "camera": {
+  "mode": "fun-track",
+  "az_deg": -95,
+  "el_deg": 22.0,
+  "fovy_deg": 34,
+  "fill": 0.8,
+  "orbit_deg": 10,
+  "distance_m": [
+   6.327,
+   9.038
+  ]
+ },
+ "revision": "soccer-2v2-11+core-0.2 (header source_sha f658ab0d)",
+ "holds": {
+  "start_s": 0.5,
+  "end_s": 1.5,
+  "note": "first/last frame held, no extra simulation"
+ },
+ "visual_changes": [
+  "daylight sky and a plain ground plane in the scene's ground colour (drawn-only)",
+  "the ball drawn white (orange in the match)",
+  "display cord nets in place of the match's box goal nets"
+ ],
+ "poster": "plan frame 458: the key moment of the clip",
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, 30 fps constant, +faststart, no audio, no text or overlays",
+ "created": "2026-10-19",
+ "contributor": "EAO team (Chi Gui), Embodied Agent Olympics"
+}

masters/roboolympics-match/soccer_jointgpu_i01/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:83ceccea78f0008cccdb9aef349293a8e2f0ae224fcd5b12664e117255573266
+size 14193706

masters/roboolympics-match/soccer_match/meta.json

@@ -0,0 +1,47 @@
+{
+ "suite": "roboolympics-match",
+ "task": "soccer_match",
+ "title": "Humanoid soccer 2 v 2, Booster T1",
+ "source": "agent-success:codex-gpt6_astra-high-roboolympics-match-soccer-match-i00-standard-realtime/roboolympics-soccer-match-i00-standard-realtime__zhEqWGZ",
+ "success_check": true,
+ "success_check_how": "the replay of the recorded run with the task's own simulator code reproduces the run's final state hash and score exactly, and the run's verdict is a win (a counted success, verifier reward 1)",
+ "agent": "GPT-6 Astra (Codex, high)",
+ "result": "Astra wins 9:0 against the built-in level-2 team (low-level actions)",
+ "trial_choice": "a counted success; the exam maintainer's pick for this task",
+ "renderer": "Blender 4.2.23 LTS Cycles path tracing (GPU OptiX), adaptive sampling (threshold 0.02, up to 128 spp), OpenImageDenoise; kinematic playback of the recorded simulation poses (no physics in Blender; the match's own geometry and colours, full-detail robot maker meshes on the same link frames); daylight: a sun lamp (2 degree disc) and a Nishita sky (no sun disc), a plain ground plane in the scene's ground colour; AgX view transform, Medium High Contrast look; motion blur 180 degrees; a tracking camera (fixed elevation, per-frame fit of the robot(s) and the ball, smoothed in time) with a slow eased 10-degree orbit, one continuous shot (the roboolympics-fun suite's look)",
+ "rendering": "path-tracing",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "frames": 386,
+ "seconds": 12.87,
+ "speed": 1.0,
+ "clip": "GPT-6 Astra's team (blue) vs the built-in level-2 team (red), match time 228.3-239.1 s: A1 works the ball forward and scores from about 2.5 m (2:0 at 239.14 s); Astra won 9:0",
+ "camera": {
+  "mode": "fun-track",
+  "az_deg": -95,
+  "el_deg": 22.0,
+  "fovy_deg": 34,
+  "fill": 0.8,
+  "orbit_deg": 10,
+  "distance_m": [
+   5.725,
+   8.778
+  ]
+ },
+ "revision": "soccer-2v2-10+core-0.2 (header source_sha 2f14b309)",
+ "holds": {
+  "start_s": 0.5,
+  "end_s": 1.5,
+  "note": "first/last frame held, no extra simulation"
+ },
+ "visual_changes": [
+  "daylight sky and a plain ground plane in the scene's ground colour (drawn-only)",
+  "the ball drawn white (orange in the match)",
+  "display cord nets in place of the match's box goal nets"
+ ],
+ "poster": "plan frame 319: the key moment of the clip",
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, 30 fps constant, +faststart, no audio, no text or overlays",
+ "created": "2026-10-19",
+ "contributor": "EAO team (Chi Gui), Embodied Agent Olympics"
+}

masters/roboolympics-match/soccer_match/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8c5b0cdd6bb2cd5cc43eaac65392d64a15b6f178c269c70a193203934ea3e015
+size 9193831

masters/roboolympics-match/soccerssl_match/meta.json

@@ -0,0 +1,46 @@
+{
+ "suite": "roboolympics-match",
+ "task": "soccerssl_match",
+ "title": "Wheeled robot soccer 5 v 5 (instance 0)",
+ "source": "oracle",
+ "success_check": true,
+ "success_check_how": "the reference solution's match, recorded with the task's own simulator code and configuration, replayed in a fresh process to the same final state hash; it wins the match",
+ "result": "the reference team wins 10:3 against the built-in level-2 team (early end at 10 goals)",
+ "why_oracle": "no evaluated agent won this task; the task's reference solution (oracle) is shown instead",
+ "oracle_recording": "played by the task's reference player with the task's own simulator code and configuration (seed, level, action level); recorded in-process (no network, no wall clock), the same player the task's oracle runs over its HTTP interface",
+ "renderer": "Blender 4.2.23 LTS Cycles path tracing (GPU OptiX), adaptive sampling (threshold 0.02, up to 128 spp), OpenImageDenoise; kinematic playback of the recorded simulation poses (no physics in Blender; the match's own geometry and colours, full-detail robot maker meshes on the same link frames); daylight: a sun lamp (2 degree disc) and a Nishita sky (no sun disc), a plain ground plane in the scene's ground colour; AgX view transform, Medium High Contrast look; motion blur 180 degrees; a tracking camera (fixed elevation, per-frame fit of the robot(s) and the ball, smoothed in time) with a slow eased 10-degree orbit, one continuous shot (the roboolympics-fun suite's look)",
+ "rendering": "path-tracing",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "frames": 325,
+ "seconds": 10.83,
+ "speed": 1.0,
+ "clip": "The task's reference solution (oracle, blue) against the built-in level-2 team: a free kick and blue4's goal for 10:3, which ends the match",
+ "camera": {
+  "mode": "fun-track",
+  "az_deg": -90,
+  "el_deg": 62.0,
+  "fovy_deg": 34,
+  "fill": 0.85,
+  "orbit_deg": 10,
+  "distance_m": [
+   3.078,
+   4.594
+  ]
+ },
+ "revision": "soccer_ssl-1 (world soccer_ssl-world-1)",
+ "holds": {
+  "start_s": 0.5,
+  "end_s": 1.5,
+  "note": "first/last frame held, no extra simulation"
+ },
+ "visual_changes": [
+  "daylight sky and a plain ground plane in the scene's ground colour (drawn-only)",
+  "the ground plane outside the field drawn in the field's carpet green (no grey band at the frame edge)"
+ ],
+ "poster": "plan frame 264: the key moment of the clip",
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, 30 fps constant, +faststart, no audio, no text or overlays",
+ "created": "2026-10-19",
+ "contributor": "EAO team (Chi Gui), Embodied Agent Olympics"
+}

masters/roboolympics-match/soccerssl_match/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ad95a093cc5f479674be887a8f69ecb47825fe63b080c9726358b5ac742abadc
+size 7386529

masters/roboolympics-match/soccerssl_match_i01/meta.json

@@ -0,0 +1,45 @@
+{
+ "suite": "roboolympics-match",
+ "task": "soccerssl_match_i01",
+ "title": "Wheeled robot soccer 5 v 5 (instance 1)",
+ "source": "scene-orbit",
+ "shows": "oracle-attempt (lost): the task's reference solution, the reference team loses 2:10 to the built-in level-4 team",
+ "why_no_video": "No evaluated agent has won this task (best: every evaluated agent lost), and the task's reference solution does not win it either. As for the roboolympics-fun suite's unsolved tasks, this video is that reference solution's match replayed exactly (not a solved run), showing its best moment: the reference team loses 2:10 to the built-in level-4 team.",
+ "success_check": false,
+ "success_check_how": "the match's own result at the end of the replay (a loss), equal to the recording",
+ "oracle_recording": "played by the task's reference player with the task's own simulator code and configuration (seed, level, action level); recorded in-process (no network, no wall clock), the same player the task's oracle runs over its HTTP interface",
+ "renderer": "Blender 4.2.23 LTS Cycles path tracing (GPU OptiX), adaptive sampling (threshold 0.02, up to 128 spp), OpenImageDenoise; kinematic playback of the recorded simulation poses (no physics in Blender; the match's own geometry and colours, full-detail robot maker meshes on the same link frames); daylight: a sun lamp (2 degree disc) and a Nishita sky (no sun disc), a plain ground plane in the scene's ground colour; AgX view transform, Medium High Contrast look; motion blur 180 degrees; a tracking camera (fixed elevation, per-frame fit of the robot(s) and the ball, smoothed in time) with a slow eased 10-degree orbit, one continuous shot (the roboolympics-fun suite's look)",
+ "rendering": "path-tracing",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "frames": 316,
+ "seconds": 10.53,
+ "speed": 1.0,
+ "clip": "The task's reference solution (oracle, blue) against the built-in level-4 team: a corner kick and blue3's goal for 2:1, the oracle's best moment in a match it loses 2:10",
+ "camera": {
+  "mode": "fun-track",
+  "az_deg": -90,
+  "el_deg": 50.0,
+  "fovy_deg": 34,
+  "fill": 0.85,
+  "orbit_deg": 10,
+  "distance_m": [
+   3.106,
+   4.163
+  ]
+ },
+ "revision": "soccer_ssl-1 (world soccer_ssl-world-1)",
+ "holds": {
+  "start_s": 0.5,
+  "end_s": 1.5,
+  "note": "first/last frame held, no extra simulation"
+ },
+ "visual_changes": [
+  "daylight sky and a plain ground plane in the scene's ground colour (drawn-only)"
+ ],
+ "poster": "plan frame 224: the key moment of the clip",
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, 30 fps constant, +faststart, no audio, no text or overlays",
+ "created": "2026-10-19",
+ "contributor": "EAO team (Chi Gui), Embodied Agent Olympics"
+}

masters/roboolympics-match/soccerssl_match_i01/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ce8e0e608d037aefb836840f7cbcce4df823b3070d744814e41fa6e56b3734b9
+size 6628352

masters/roboolympics-match/tabletennis_duelrl_i01/meta.json

@@ -0,0 +1,45 @@
+{
+ "suite": "roboolympics-match",
+ "task": "tabletennis_duelrl_i01",
+ "title": "Table tennis duel vs an RL opponent, two Franka arms",
+ "source": "oracle",
+ "success_check": true,
+ "success_check_how": "the reference solution's match, recorded with the task's own simulator code and configuration, replayed in a fresh process to the same final state hash; it wins the match",
+ "result": "the reference solution wins 5:1 against the task's RL opponent (seat B, a second Franka arm)",
+ "why_oracle": "no evaluated agent won this task; the task's reference solution (oracle) is shown instead",
+ "oracle_recording": "played by the task's reference player with the task's own simulator code and configuration (seed, level, action level); recorded in-process (no network, no wall clock), the same player the task's oracle runs over its HTTP interface",
+ "renderer": "Blender 4.2.23 LTS Cycles path tracing (GPU OptiX), adaptive sampling (threshold 0.02, up to 128 spp), OpenImageDenoise; kinematic playback of the recorded simulation poses (no physics in Blender; the match's own geometry and colours, full-detail robot maker meshes on the same link frames); daylight: a sun lamp (2 degree disc) and a Nishita sky (no sun disc), a plain ground plane in the scene's ground colour; AgX view transform, Medium High Contrast look; motion blur 180 degrees; a tracking camera (fixed elevation, per-frame fit of the robot(s) and the ball, smoothed in time) with a slow eased 10-degree orbit, one continuous shot (the roboolympics-fun suite's look)",
+ "rendering": "path-tracing",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "frames": 344,
+ "seconds": 11.47,
+ "speed": 1.0,
+ "clip": "Reference solution (oracle, seat A, near arm) against the task's built-in RL opponent (far arm): the last three points, all won by the oracle, ending the match 5:1; the ball is re-placed for each serve",
+ "camera": {
+  "mode": "fun-track",
+  "az_deg": -150,
+  "el_deg": 22.0,
+  "fovy_deg": 34,
+  "fill": 1.0,
+  "orbit_deg": 10,
+  "distance_m": [
+   6.335,
+   6.412
+  ]
+ },
+ "revision": "wallclock-15+duel-wallclock-12+low-inloop-1+core-0.2",
+ "holds": {
+  "start_s": 0.5,
+  "end_s": 1.5,
+  "note": "first/last frame held, no extra simulation"
+ },
+ "visual_changes": [
+  "daylight sky and a plain ground plane in the scene's ground colour (drawn-only)"
+ ],
+ "poster": "plan frame 66: the key moment of the clip",
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, 30 fps constant, +faststart, no audio, no text or overlays",
+ "created": "2026-10-19",
+ "contributor": "EAO team (Chi Gui), Embodied Agent Olympics"
+}

masters/roboolympics-match/tabletennis_duelrl_i01/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:33fd584d44b9d21062bf932b65ad809483a7b1c3c9e2699219089f03e5dd5397
+size 5173169

masters/roboolympics-match/tabletennis_match_i01/meta.json

@@ -0,0 +1,49 @@
+{
+ "suite": "roboolympics-match",
+ "task": "tabletennis_match_i01",
+ "title": "Table tennis match, Franka arm (instance 1)",
+ "source": "agent-success:codex-gpt6_astra-high-roboolympics-match-tabletennis-match-i01-standard-realtime/roboolympics-tabletennis-match-i__Xz9mx4f",
+ "success_check": true,
+ "success_check_how": "the replay of the recorded run with the task's own simulator code reproduces the run's final state hash and score exactly, and the run's verdict is a win (a counted success, verifier reward 1)",
+ "agent": "GPT-6 Astra (Codex, high)",
+ "result": "Astra wins 6:4 against the built-in level-2 opponent",
+ "trial_choice": "a counted success; the exam maintainer's pick for this task",
+ "renderer": "Blender 4.2.23 LTS Cycles path tracing (GPU OptiX), adaptive sampling (threshold 0.02, up to 128 spp), OpenImageDenoise; kinematic playback of the recorded simulation poses (no physics in Blender; the match's own geometry and colours, full-detail robot maker meshes on the same link frames); daylight: a sun lamp (2 degree disc) and a Nishita sky (no sun disc), a plain ground plane in the scene's ground colour; AgX view transform, Medium High Contrast look; motion blur 180 degrees; a tracking camera (fixed elevation, per-frame fit of the robot(s) and the ball, smoothed in time) with a slow eased 10-degree orbit, one continuous shot (the roboolympics-fun suite's look)",
+ "rendering": "path-tracing",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "frames": 418,
+ "seconds": 13.93,
+ "speed": 1.0,
+ "clip": "GPT-6 Astra (Franka Panda on an x-y rail) against the built-in level-2 opponent: the 3:3 and 4:3 points, including the longest rally of the match, which Astra wins with a return too wide for the opponent (Astra wins 6:4); the opponent arm (a Franka Panda like the player's) is drawn for display only, posed to meet each recorded return; the ball is re-placed for each serve",
+ "camera": {
+  "mode": "fun-track",
+  "az_deg": -150,
+  "el_deg": 22.0,
+  "fovy_deg": 34,
+  "fill": 1.0,
+  "orbit_deg": 10,
+  "distance_m": [
+   6.326,
+   6.42
+  ]
+ },
+ "revision": "wallclock-15+duel-wallclock-12+low-inloop-1+core-0.2",
+ "holds": {
+  "start_s": 0.5,
+  "end_s": 1.5,
+  "note": "first/last frame held, no extra simulation"
+ },
+ "visual_changes": [
+  "daylight sky and a plain ground plane in the scene's ground colour (drawn-only)",
+  "an opponent arm (the same arm and racket as the player's, mirrored on the far side) drawn for display only: posed by kinematics (inverse kinematics to each recorded return of the scripted opponent, swing and ready poses between them); the ball path is the recording's; in the task the scripted opponent has no body"
+ ],
+ "notes": [
+  "opponent arm drawn for display only; in the task the scripted opponent has no body"
+ ],
+ "poster": "plan frame 264: the key moment of the clip",
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, 30 fps constant, +faststart, no audio, no text or overlays",
+ "created": "2026-10-19",
+ "contributor": "EAO team (Chi Gui), Embodied Agent Olympics"
+}

masters/roboolympics-match/tabletennis_match_i01/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b8e1c83ccfaf6061e3167c9f4049e756e79e0c4f07f92a7ef0ea8d695893788a
+size 5957540

masters/roboolympics-match/tabletennis_match_i02/meta.json

@@ -0,0 +1,50 @@
+{
+ "suite": "roboolympics-match",
+ "task": "tabletennis_match_i02",
+ "title": "Table tennis match, Franka arm (instance 2)",
+ "source": "scene-orbit",
+ "shows": "oracle-attempt (lost): the task's reference solution: it loses 3:5 to the built-in level-3 opponent",
+ "why_no_video": "No evaluated agent has won this task (best: GPT-6 Astra 1:5), and the task's reference solution does not win it either. As for the roboolympics-fun suite's unsolved tasks, this video is that reference solution's match replayed exactly (not a solved run), showing its best moment: the reference solution loses 3:5 to the built-in level-3 opponent.",
+ "success_check": false,
+ "success_check_how": "the match's own result at the end of the replay (a loss), equal to the recording",
+ "oracle_recording": "played by the task's reference player with the task's own simulator code and configuration (seed, level, action level); recorded in-process (no network, no wall clock), the same player the task's oracle runs over its HTTP interface",
+ "renderer": "Blender 4.2.23 LTS Cycles path tracing (GPU OptiX), adaptive sampling (threshold 0.02, up to 128 spp), OpenImageDenoise; kinematic playback of the recorded simulation poses (no physics in Blender; the match's own geometry and colours, full-detail robot maker meshes on the same link frames); daylight: a sun lamp (2 degree disc) and a Nishita sky (no sun disc), a plain ground plane in the scene's ground colour; AgX view transform, Medium High Contrast look; motion blur 180 degrees; a tracking camera (fixed elevation, per-frame fit of the robot(s) and the ball, smoothed in time) with a slow eased 10-degree orbit, one continuous shot (the roboolympics-fun suite's look)",
+ "rendering": "path-tracing",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "frames": 319,
+ "seconds": 10.63,
+ "speed": 1.0,
+ "clip": "Reference solution (oracle) against the built-in level-3 opponent: a lost point, then a point the oracle wins with a return too wide for the opponent (the oracle loses the match 3:5); the opponent arm (a Franka Panda like the player's) is drawn for display only, posed to meet each recorded return; the ball is re-placed for each serve",
+ "camera": {
+  "mode": "fun-track",
+  "az_deg": -150,
+  "el_deg": 22.0,
+  "fovy_deg": 34,
+  "fill": 1.0,
+  "orbit_deg": 10,
+  "distance_m": [
+   6.322,
+   6.511
+  ]
+ },
+ "revision": "wallclock-15+duel-wallclock-12+low-inloop-1+core-0.2",
+ "holds": {
+  "start_s": 0.5,
+  "end_s": 1.5,
+  "note": "first/last frame held, no extra simulation"
+ },
+ "visual_changes": [
+  "daylight sky and a plain ground plane in the scene's ground colour (drawn-only)",
+  "an opponent arm (the same arm and racket as the player's, mirrored on the far side) drawn for display only: posed by kinematics (inverse kinematics to each recorded return of the scripted opponent, swing and ready poses between them); the ball path is the recording's; in the task the scripted opponent has no body"
+ ],
+ "notes": [
+  "opponent arm drawn for display only; in the task the scripted opponent has no body",
+  "the task's own documentation reports a live oracle run that wins this instance; this recording of the same reference player (in-process, no network or wall clock) loses it, so it is labelled as a lost attempt"
+ ],
+ "poster": "plan frame 215: the key moment of the clip",
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, 30 fps constant, +faststart, no audio, no text or overlays",
+ "created": "2026-10-19",
+ "contributor": "EAO team (Chi Gui), Embodied Agent Olympics"
+}

masters/roboolympics-match/tabletennis_match_i02/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3ca68eeabe450fcc3b4c624d6b28ace467452998ecb1b8e57139b79dbb143377
+size 4899357

masters/roboolympics-match/tabletennis_match_i03/meta.json

@@ -0,0 +1,50 @@
+{
+ "suite": "roboolympics-match",
+ "task": "tabletennis_match_i03",
+ "title": "Table tennis match, Franka arm (instance 3)",
+ "source": "scene-orbit",
+ "shows": "oracle-attempt (lost): the task's reference solution: it loses 3:5 to the built-in level-4 (top tier) opponent",
+ "why_no_video": "No evaluated agent has won this task (best: Claude Opus 5.5 2:5), and the task's reference solution does not win it either. As for the roboolympics-fun suite's unsolved tasks, this video is that reference solution's match replayed exactly (not a solved run), showing its best moment: the reference solution loses 3:5 to the built-in level-4 (top tier) opponent.",
+ "success_check": false,
+ "success_check_how": "the match's own result at the end of the replay (a loss), equal to the recording",
+ "oracle_recording": "played by the task's reference player with the task's own simulator code and configuration (seed, level, action level); recorded in-process (no network, no wall clock), the same player the task's oracle runs over its HTTP interface",
+ "renderer": "Blender 4.2.23 LTS Cycles path tracing (GPU OptiX), adaptive sampling (threshold 0.02, up to 128 spp), OpenImageDenoise; kinematic playback of the recorded simulation poses (no physics in Blender; the match's own geometry and colours, full-detail robot maker meshes on the same link frames); daylight: a sun lamp (2 degree disc) and a Nishita sky (no sun disc), a plain ground plane in the scene's ground colour; AgX view transform, Medium High Contrast look; motion blur 180 degrees; a tracking camera (fixed elevation, per-frame fit of the robot(s) and the ball, smoothed in time) with a slow eased 10-degree orbit, one continuous shot (the roboolympics-fun suite's look)",
+ "rendering": "path-tracing",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "frames": 343,
+ "seconds": 11.43,
+ "speed": 1.0,
+ "clip": "Reference solution (oracle) against the built-in level-4 top-tier opponent: two lost points, then a serve the opponent cannot return legally (the oracle loses the match 3:5); the opponent arm (a Franka Panda like the player's) is drawn for display only, posed to meet each recorded return; the ball is re-placed for each serve",
+ "camera": {
+  "mode": "fun-track",
+  "az_deg": -150,
+  "el_deg": 22.0,
+  "fovy_deg": 34,
+  "fill": 1.0,
+  "orbit_deg": 10,
+  "distance_m": [
+   6.316,
+   6.411
+  ]
+ },
+ "revision": "wallclock-15+duel-wallclock-12+low-inloop-1+core-0.2",
+ "holds": {
+  "start_s": 0.5,
+  "end_s": 1.5,
+  "note": "first/last frame held, no extra simulation"
+ },
+ "visual_changes": [
+  "daylight sky and a plain ground plane in the scene's ground colour (drawn-only)",
+  "an opponent arm (the same arm and racket as the player's, mirrored on the far side) drawn for display only: posed by kinematics (inverse kinematics to each recorded return of the scripted opponent, swing and ready poses between them); the ball path is the recording's; in the task the scripted opponent has no body"
+ ],
+ "notes": [
+  "opponent arm drawn for display only; in the task the scripted opponent has no body",
+  "the task's own documentation reports a live oracle run that wins this instance; this recording of the same reference player (in-process, no network or wall clock) loses it, so it is labelled as a lost attempt"
+ ],
+ "poster": "plan frame 234: the key moment of the clip",
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, 30 fps constant, +faststart, no audio, no text or overlays",
+ "created": "2026-10-19",
+ "contributor": "EAO team (Chi Gui), Embodied Agent Olympics"
+}