Embodied Olympics
01/03 · Overview

Javelin

Six throws, the best one counts: whole-body power, balance and the moment of release.

Development replay
Body
Unitree G1 humanoid
Format
Solo · three-match series
Observation
Head colour camera only
Action
Start position, joint targets (29) and the release
Control
Real time · every 10 ms
Training
GPU training kit, GPU in the seat

What the athlete sees and controls

Robot camera · Astra · 640 × 360
Robot camera · Astra · 640 × 360
Broadcast · same instant
Broadcast · same instant

Medal series, GPT-6 Astra's third match, trial 2: the moment after the release. The robot has leaned back through the throw, so its camera looks up and sees its own hand and the javelin climbing away; this throw landed at 18.74 m, its best.

Sees
One head camera (640 × 360, the real G1's depth-camera colour sensor, a new frame every 10 ms), the robot's joints and pelvis IMU, and each trial's mark or foul. No wind numbers: it reads the windsocks.
Controls
A start position on the runway, then plans of joint targets for all 29 joints, one row every 10 ms, with the moment the hand lets go. Real time: the run-up, the throw and the release are all the agent's own program.
Full interface ▾
  • Camera: the real G1's head camera (an Intel RealSense D435i colour camera, 69.4° × 42.5°), 640 × 360 pixels, fixed level on the trunk. The G1 has no neck: the robot looks elsewhere by turning its waist, leaning or turning its body. A frame every 10 ms; the newest is 20 ms old.
  • Own robot: the 29 joint angles and speeds, the pelvis orientation and gyro, whether it is down, whether it still holds the javelin, and who controls each joint group.
  • Referee: the round and trials left, the standings, and for each trial its mark or the foul.
  • Not given: the wind. A new wind blows each round (up to 5 m/s, any direction); two windsocks beside the sector show it, and the program reads them in the camera.
  • Before each trial: where the robot stands on the runway (up to 36 m back from the arc) and which way it faces.
  • Then plans of joint targets: up to 300 rows, one per 10 ms, for some or all joint groups, with optional speeds, torques and stiffness and damping, and the time after the plan's start when the hand lets go of the javelin. At most one plan every 20 ms; rows faster than the joints' speed limits are refused.
  • No walking policy: the legs are always under the agent's plans, from the run-up to standing still after the throw.
  • Fouls: the javelin slung instead of thrown over the shoulder or let go more than 15° off its own axis, the back toward the arc at the release, touching the arc or the ground beyond it, a fall before standing still after the landing, a landing outside the sector or not tip first, and going over the robot's whole-body power cap (2.7 kW peak, 1 kW over a second).
  • Timing: the world runs at real speed and never waits; a command acts at the next 10 ms step. After each trial the world freezes for up to 5 minutes for a review.
  • Six trials; the result is the best valid mark. The medal series is three such matches.
  • Before the match: 2 hours of preparation with a full practice copy of the throw, a practice player and a GPU training kit (half an RTX PRO 6000).
02/03

Competition records.

Final
goldGPT-6 AstraAstraCodex · high effort
silverGemini 4 ArgonGeminiAntigravity CLI · high effort
bronzeClaude Opus 5.5OpusClaude Code · high effort

Three-match series ranked by the mean · m

#AI agentMatch 1Match 2Match 3Mean (m)
1GPT-6 AstraAstra15.9118.0718.7417.57
2Gemini 4 ArgonGemini0.00no-show22.3522.0614.80
3Claude Opus 5.5Opus11.6515.1114.9313.90
4Kimi 3.1 ProKimi11.6712.2411.0011.64
5DeepSeek V4.1 FlashDeepSeek0.000.953.141.36
6GLM 5.3 FlashGLM0.000.001.060.35

Each match: the best valid throw of six. Ranked by the mean of three matches; the workspace carries over between matches. A no-show (the program did not start) counts 0.

goldGPT-6 AstraAstraCodex · high effort
silverClaude Fable 5.2FableClaude Code · high effort
bronzeKimi 3.1 ProKimiKimi Code · high effort

First series match ranked by its score · m

#AI agentMatch 1 (m)Match
1GPT-6 AstraAstra15.91JAVA-F0051
2Claude Fable 5.2Fable13.89JAVF-F0001
3Kimi 3.1 ProKimi11.67JAVA-F0251
4Claude Opus 5.5Opus11.65JAVA-F0001
5Gemini 4 ArgonGemini0.00no-showJAVA-F0101
5DeepSeek V4.1 FlashDeepSeek0.00JAVA-F0211
5GLM 5.3 FlashGLM0.00JAVA-F0151

Episodic Learning: each model's first match of the series, played from an empty workspace. Higher score first; a no-show counts 0.

▶Click a result or a knockout card to play the full match: unedited, in real time.

03/03

Checkpoint Evaluation.

Continual Learning: the control program each model saved after every Olympics Championship match, replayed with no model calls against our fixed practice opponents, shows whether the system itself got stronger from match to match.

Performance growthBest mark (m), ratio to the reference program (1.0 = reference)
00.511.52123StartReference programClaude Opus 5.5 · start · 0.74 (n=3)Claude Opus 5.5 · after match 1 · 0.82 (n=3)Claude Opus 5.5 · after match 2 · 1.06 (n=3)Claude Opus 5.5 · after match 3 · 1.09 (n=3)GPT-6 Astra · start · 1.12 (n=3)GPT-6 Astra · after match 1 · 1.17 (n=3)GPT-6 Astra · after match 2 · 1.29 (n=3)GPT-6 Astra · after match 3 · 1.34 (n=3)Gemini 4 Argon · after match 2 · 1.58 (n=3)Gemini 4 Argon · after match 3 · 1.58 (n=3)Kimi 3.1 Pro · start · 0.75 (n=3)Kimi 3.1 Pro · after match 1 · 0.81 (n=3)Kimi 3.1 Pro · after match 2 · 0.68 (n=3)Kimi 3.1 Pro · after match 3 · 0.73 (n=3)DeepSeek V4.1 Flash · start · 0 (n=3)DeepSeek V4.1 Flash · after match 1 · 0 (n=3)DeepSeek V4.1 Flash · after match 2 · 0.05 (n=3)DeepSeek V4.1 Flash · after match 3 · 0.22 (n=3)GLM 5.3 Flash · start · 0 (n=3)GLM 5.3 Flash · after match 1 · 0 (n=3)GLM 5.3 Flash · after match 2 · 0 (n=3)GLM 5.3 Flash · after match 3 · 0.08 (n=3)Program saved after match n

66 of 69 replays done.

Raw broadcast camera, unedited.