EO Match: table tennis videos (G1 GPU tier, joint-level match, RL duel) from agents' wins

masters/roboolympics-match/tabletennis_duelrl_i01/meta.json

@@ -2,44 +2,83 @@
  "suite": "roboolympics-match",
  "task": "tabletennis_duelrl_i01",
  "title": "Table tennis duel vs an RL opponent, two Franka arms",
- "source": "oracle",
+ "task_dir": "embodied-olympics-match-tabletennis-duelrl-i01-privileged",
+ "source": "agent-success:claude_code-opus_5_5-high-embodied-olympics-match-tabletennis-duelrl-i01-privileged/embodied-olympics-match-tableten__PxbxC7g",
  "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)",
+ "success_check_how": "privileged trial: the verifier played the agent's program in a fresh match and recorded the episode; that episode was replayed with the sport's own core replay (core.episode.replay, the verifier's fresh-process replay check) in the task's verifier simulator; the final state hash (994d6b1f90e5583e075d3c110e31a576c0a309319dcb4bf51711f0f556fa8e42), the score, the winner and the step count equal the verifier's replay check, and the verdict is a win (a counted success, reward 1)",
+ "agent": "Claude Opus 5.5 (Claude Code, high)",
+ "run": "claude_code-opus_5_5-high",
+ "mode": "privileged",
+ "trial": "claude_code-opus_5_5-high-embodied-olympics-match-tabletennis-duelrl-i01-privileged/embodied-olympics-match-tableten__PxbxC7g",
+ "result": "Opus's program wins 5:3 against the task's RL opponent (seat B, a second Franka arm)",
+ "trial_choice": "no agent has won this task in standard or standard-realtime mode, and GPT-6 Astra did not win it in privileged mode; Claude Opus 5.5's privileged success (its program, played by the verifier in a fresh match) is shown",
+ "renderer": "Blender 4.5.3 LTS Cycles path tracing (GPU OptiX), adaptive sampling (threshold 0.01, 64-128 spp), OpenImageDenoise with albedo + normal passes; kinematic playback of the exact replayed simulation poses (no physics in Blender; the match's spectator scene (the sport's presentation scene: MuJoCo Menagerie's full Franka Panda meshes on the same link frames, same kinematics)); daylight: a sun lamp (2 degree disc) and a Nishita sky (no sun disc), a plain ground plane in light grey (the suite's Franka table tennis look); AgX view transform, Medium High Contrast look; fixed 3/4 camera fitted to the robots and the table over the clip, with a slow eased 10-degree orbit, one continuous shot; native 2560x1440",
  "rendering": "path-tracing",
+ "media_standard": "v2 (2026-10-16)",
  "width": 2560,
  "height": 1440,
  "fps": 30,
- "frames": 344,
- "seconds": 11.47,
+ "frames": 540,
+ "seconds": 18.0,
  "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",
+ "clip": "Claude Opus 5.5's program (near arm, seat A) against the task's RL opponent (far arm), from 2:2 to the end: 3:2 (the ball hits the opponent's arm), 3:3 (Opus misses), 4:3 (the opponent's serve fault), then a five-hit rally won when the ball hits the opponent's arm, 5:3",
+ "clip_steps": {
+  "k0": 1640,
+  "k1": 3237,
+  "sim_s": 15.97,
+  "control_hz": 100.0,
+  "score_before": {
+   "A": 2,
+   "B": 2
+  }
+ },
  "camera": {
-  "mode": "fun-track",
-  "az_deg": -150,
+  "mode": "fixed",
+  "frames_on": "robots,table",
+  "az_deg": -150.0,
   "el_deg": 22.0,
-  "fovy_deg": 34,
-  "fill": 1.0,
-  "orbit_deg": 10,
-  "distance_m": [
-   6.335,
-   6.412
-  ]
+  "fovy_deg": 34.0,
+  "fill": 0.8,
+  "target": [
+   -0.0122,
+   -0.1901,
+   -0.1942
+  ],
+  "distance_m": 4.8353,
+  "orbit_deg": 10.0
  },
  "revision": "wallclock-15+duel-wallclock-12+low-inloop-1+core-0.2",
+ "replay": {
+  "final_state_sha": "994d6b1f90e5583e075d3c110e31a576c0a309319dcb4bf51711f0f556fa8e42",
+  "expected_from": "replay_check.json (the verifier's replay)",
+  "expected_state_sha": "994d6b1f90e5583e075d3c110e31a576c0a309319dcb4bf51711f0f556fa8e42",
+  "scores": {
+   "A": 5,
+   "B": 3
+  },
+  "winner": "A",
+  "steps": 3237,
+  "checks": {
+   "state_sha": true,
+   "winner": true,
+   "scores": true,
+   "k": true
+  },
+  "scene": "presentation",
+  "episode_sha256": "5c0a26f785c8cfcc06276ea0784438310dd7a5d23a8b921de7216d6800352528",
+  "mujoco": "3.3.7"
+ },
  "holds": {
   "start_s": 0.5,
   "end_s": 1.5,
-  "note": "first/last frame held, no extra simulation"
+  "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)"
+  "daylight sky and a plain ground plane instead of the venue's floor (drawn only)"
  ],
- "poster": "plan frame 66: the key moment of the clip",
+ "poster": "last frame of the clip (just after the clip's last point)",
  "encode": "libx264 preset slow CRF 16, yuv420p, bt709, 30 fps constant, +faststart, no audio, no text or overlays",
- "created": "2026-10-19",
+ "seconds_per_frame": 2.16,
+ "created": "2026-11-03",
  "contributor": "EAO team (Chi Gui), Embodied Agent Olympics"
 }

masters/roboolympics-match/tabletennis_duelrl_i01/video.mp4

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:33fd584d44b9d21062bf932b65ad809483a7b1c3c9e2699219089f03e5dd5397
-size 5173169
+oid sha256:93fcc2c338da39503e80c9c235922a240d78aa6eadb3d2365535543274424bce
+size 7690304

masters/roboolympics-match/tabletennis_matchlow_i01/meta.json

@@ -1,49 +1,85 @@
 {
  "suite": "roboolympics-match",
  "task": "tabletennis_matchlow_i01",
- "title": "Table tennis match, low-level control (instance 1)",
- "source": "scene-orbit",
- "shows": "oracle-attempt (lost): the task's reference solution: it (its hit skill in the loop) loses 0:5 to the built-in level-1 opponent",
- "why_no_video": "No evaluated agent has won this task (best: GPT-6 Astra 3:5), 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 (its hit skill in the loop) loses 0:5 to the built-in level-1 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)",
+ "title": "Table tennis match, joint-level control, Franka arm",
+ "task_dir": "embodied-olympics-match-tabletennis-matchlow-i01-privileged",
+ "source": "agent-success:codex-gpt6_astra-high-embodied-olympics-match-tabletennis-matchlow-i01-privileged/embodied-olympics-match-tableten__NRVBmjg",
+ "success_check": true,
+ "success_check_how": "privileged trial: the verifier played the agent's program in a fresh match and recorded the episode; that episode was replayed with the sport's own core replay (core.episode.replay, the verifier's fresh-process replay check) in the task's verifier simulator; the final state hash (09acaafa661d1e15be80120fac2267d09a28c3d186aacd3bdfb680cde4aa0bae), the score, the winner and the step count equal the verifier's replay check, and the verdict is a win (a counted success, reward 1)",
+ "agent": "GPT-6 Astra (Codex, high)",
+ "run": "codex-gpt6_astra-high",
+ "mode": "privileged",
+ "trial": "codex-gpt6_astra-high-embodied-olympics-match-tabletennis-matchlow-i01-privileged/embodied-olympics-match-tableten__NRVBmjg",
+ "result": "Astra's program wins 7:6 against the built-in opponent",
+ "trial_choice": "no agent has won this task in standard or standard-realtime mode; GPT-6 Astra's privileged success (its program, played by the verifier in a fresh match) is shown",
+ "renderer": "Blender 4.5.3 LTS Cycles path tracing (GPU OptiX), adaptive sampling (threshold 0.01, 64-128 spp), OpenImageDenoise with albedo + normal passes; kinematic playback of the exact replayed simulation poses (no physics in Blender; the match's spectator scene (the sport's presentation scene: MuJoCo Menagerie's full Franka Panda meshes on the same link frames, same kinematics)); daylight: a sun lamp (2 degree disc) and a Nishita sky (no sun disc), a plain ground plane in light grey (the suite's Franka table tennis look); AgX view transform, Medium High Contrast look; fixed 3/4 camera fitted to the robots and the table over the clip, with a slow eased 10-degree orbit, one continuous shot; native 2560x1440",
  "rendering": "path-tracing",
+ "media_standard": "v2 (2026-10-16)",
  "width": 2560,
  "height": 1440,
  "fps": 30,
- "frames": 370,
- "seconds": 12.33,
+ "frames": 511,
+ "seconds": 17.03,
  "speed": 1.0,
- "clip": "Reference solution (oracle) at the low level (its own hit skill in the loop) against the built-in level-1 opponent: the oracle swings at two serves and misses, then drops its own serve (it loses 0:5, as the task README documents for this seed); 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",
+ "clip": "GPT-6 Astra's joint-level program (Franka Panda on an x-y rail) against the built-in opponent, the end of a close match: from 5:4, Astra misses (5:5), wins with a return too wide for the opponent (6:5), misses again (6:6), then wins the match with another return too wide for the opponent (7:6)",
+ "clip_steps": {
+  "k0": 3589,
+  "k1": 5090,
+  "sim_s": 15.01,
+  "control_hz": 100.0,
+  "score_before": [
+   5,
+   4
+  ]
+ },
  "camera": {
-  "mode": "fun-track",
-  "az_deg": -150,
+  "mode": "fixed",
+  "frames_on": "robots,table",
+  "az_deg": -150.0,
   "el_deg": 22.0,
-  "fovy_deg": 34,
-  "fill": 1.0,
-  "orbit_deg": 10,
-  "distance_m": [
-   6.319,
-   6.392
-  ]
+  "fovy_deg": 34.0,
+  "fill": 0.8,
+  "target": [
+   -0.1134,
+   -0.5204,
+   -0.1774
+  ],
+  "distance_m": 4.2258,
+  "orbit_deg": 10.0
  },
  "revision": "wallclock-15+duel-wallclock-12+low-inloop-1+core-0.2",
+ "replay": {
+  "final_state_sha": "09acaafa661d1e15be80120fac2267d09a28c3d186aacd3bdfb680cde4aa0bae",
+  "expected_from": "replay_check.json (the verifier's replay)",
+  "expected_state_sha": "09acaafa661d1e15be80120fac2267d09a28c3d186aacd3bdfb680cde4aa0bae",
+  "scores": {
+   "A": 7,
+   "opponent": 6
+  },
+  "winner": "A",
+  "steps": 5090,
+  "checks": {
+   "state_sha": true,
+   "winner": true,
+   "scores": true,
+   "k": true
+  },
+  "scene": "presentation",
+  "episode_sha256": "e54fd95d03515b86bd3e8fa16448567e0a855493e776799c571c6e8d67f3f3d8",
+  "mujoco": "3.3.7"
+ },
  "holds": {
   "start_s": 0.5,
   "end_s": 1.5,
-  "note": "first/last frame held, no extra simulation"
+  "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"
+  "daylight sky and a plain ground plane instead of the venue's floor (drawn only)",
+  "no opponent body is drawn: in the task the scripted opponent has no body (its returns come from its virtual paddle), so the ball comes back over the net on its own"
  ],
- "poster": "plan frame 86: the key moment of the clip",
+ "poster": "last frame of the clip (just after the clip's last point)",
  "encode": "libx264 preset slow CRF 16, yuv420p, bt709, 30 fps constant, +faststart, no audio, no text or overlays",
- "created": "2026-10-19",
+ "seconds_per_frame": 2.68,
+ "created": "2026-11-03",
  "contributor": "EAO team (Chi Gui), Embodied Agent Olympics"
 }

masters/roboolympics-match/tabletennis_matchlow_i01/video.mp4

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f3bd67c9791bc32e17992bf1985282b8d8b3d738de60b34c4e82cb6c90110f95
-size 5447857
+oid sha256:3aefa2e63329b3a880f7464642ff893c41a91369fa70393285207cd525c9eaf8
+size 5902322

masters/roboolympics-match/tabletennisg1_matchgpu_i01/meta.json

@@ -0,0 +1,80 @@
+{
+ "suite": "roboolympics-match",
+ "task": "tabletennisg1_matchgpu_i01",
+ "title": "Table tennis between two Unitree G1 humanoids (GPU tier)",
+ "task_dir": "embodied-olympics-match-tabletennisg1-matchgpu-i01-privileged",
+ "source": "agent-success:codex-gpt6_luna-xhigh-embodied-olympics-match-tabletennisg1-matchgpu-i01-privileged/embodied-olympics-match-tableten__jdBHgWK",
+ "success_check": true,
+ "success_check_how": "privileged trial: the verifier played the agent's program in a fresh match and recorded the episode; that episode was replayed with the sport's own core replay (core.episode.replay, the verifier's fresh-process replay check) in the task's verifier simulator; the final state hash (1d33d773da6fc04f5e66b39ff71c70bb445f4fa02e78c4b74c682ae1a87895b4), the score, the winner and the step count equal the verifier's replay check, and the verdict is a win (a counted success, reward 1)",
+ "agent": "GPT-6 Luna (Codex, xhigh)",
+ "run": "codex-gpt6_luna-xhigh",
+ "mode": "privileged",
+ "trial": "codex-gpt6_luna-xhigh-embodied-olympics-match-tabletennisg1-matchgpu-i01-privileged/embodied-olympics-match-tableten__jdBHgWK",
+ "result": "Luna's program wins 5:1 against the platform's G1 at level 4 (the BOSS)",
+ "trial_choice": "the only success on this task in any mode (GPT-6 Luna, privileged): the program the agent submitted, played by the verifier in a fresh match and rendered after the run",
+ "renderer": "Blender 4.5.3 LTS Cycles path tracing (GPU OptiX), adaptive sampling (threshold 0.01, 64-128 spp), OpenImageDenoise with albedo + normal passes; kinematic playback of the exact replayed simulation poses (no physics in Blender; the match's own geometry and colours (Unitree G1 visual meshes of the robot library, MuJoCo Menagerie's model)); 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; fixed 3/4 camera fitted to the robots and the table over the clip, with a slow eased 10-degree orbit, one continuous shot; native 2560x1440",
+ "rendering": "path-tracing",
+ "media_standard": "v2 (2026-10-16)",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "frames": 548,
+ "seconds": 18.27,
+ "speed": 1.0,
+ "clip": "GPT-6 Luna's program on the near G1 (seat A) against the platform's G1 at level 4: the first three points of the match, all won by Luna's G1 (the opponent's shot goes out, hits its own half, then volleys the ball), 3:0; Luna won 5:1",
+ "clip_steps": {
+  "k0": 0,
+  "k1": 1624,
+  "sim_s": 16.24,
+  "control_hz": 100.0,
+  "score_before": null
+ },
+ "camera": {
+  "mode": "fixed",
+  "frames_on": "robots,table",
+  "az_deg": -35.0,
+  "el_deg": 20.0,
+  "fovy_deg": 34.0,
+  "fill": 0.86,
+  "target": [
+   -0.1262,
+   -0.2159,
+   -0.2802
+  ],
+  "distance_m": 4.6957,
+  "orbit_deg": 10.0
+ },
+ "revision": "g1tt-10+core-0.5",
+ "replay": {
+  "final_state_sha": "1d33d773da6fc04f5e66b39ff71c70bb445f4fa02e78c4b74c682ae1a87895b4",
+  "expected_from": "replay_check.json (the verifier's replay)",
+  "expected_state_sha": "1d33d773da6fc04f5e66b39ff71c70bb445f4fa02e78c4b74c682ae1a87895b4",
+  "scores": {
+   "A": 5,
+   "opponent": 1
+  },
+  "winner": "A",
+  "steps": 3436,
+  "checks": {
+   "state_sha": true,
+   "winner": true,
+   "scores": true,
+   "k": true
+  },
+  "scene": "match",
+  "episode_sha256": "931f42bf58e9f079c045e262c279c290b0edee8af8495d089cec84af2a23508e",
+  "mujoco": "3.3.7"
+ },
+ "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 instead of the venue's floor (drawn only)"
+ ],
+ "poster": "last frame of the clip (just after the clip's last point)",
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, 30 fps constant, +faststart, no audio, no text or overlays",
+ "seconds_per_frame": 2.98,
+ "created": "2026-11-03"
+}

masters/roboolympics-match/tabletennisg1_matchgpu_i01/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:047474415ac36d16c60c339943eb9dac3778770dabd0805ec2cb90b2e8c70d87
+size 10304965