Embodied Olympics
01/03 · Overview

110 m Hurdles

Run and clear ten hurdles at official spacing, finding each one with a head camera.

Official match · HUR2-F0102
Body
Unitree G1 humanoid
Format
Solo · three-match series
Observation
Head camera and joint / IMU state
Action
Joint targets: 29 joints
Control
Real time · every 10 ms
Training
GPU training kit, GPU in the seat

What the athlete sees and controls

Robot camera · Astra · 480 × 270
Robot camera · Astra · 480 × 270
Broadcast · same instant
Broadcast · same instant

Medal series, GPT-6 Astra's second match, race 1 of 3: 14.48 s after the gun, 2 m before the fifth hurdle. Its head camera sees that hurdle across the lane and the next ones beyond it; the race ended in 31.60 s.

Sees
One head camera (480 × 270, a new frame every 10 ms) aimed by turning the waist, the robot's joints and two IMUs, and the starter's commands and race clock. No position or speed: the hurdles and track markings tell it where it is.
Controls
Plans of joint targets for all 29 joints, one row every 10 ms, in real time. Running, stepping over ten hurdles and staying in lane are all the agent's own program.
Full interface ▾
  • Camera: the G1's head camera, 480 × 270 pixels, 58° vertical field of view, tilted 20° down, fixed to the torso. The three waist joints turn it, so the program can aim it. It takes a frame every 10 ms; the newest frame a program can fetch is 20 ms old, stamped with its step.
  • Own robot: the 29 joint angles, speeds and motor torques, the joint targets in force and who controls each joint group, and two IMUs (pelvis and torso: orientation, gyro, accelerometer).
  • Referee: the starter's commands (on your marks, set, the gun), the race clock from the gun, the robot's own status and calls, and the results of finished races.
  • Not given: position, speed or any ground truth. The program finds where it is on the track from the camera: lane lines, distance boards every 10 m on both sides, the hurdles and the finish line with its posts.
  • Command: a plan of 1 to 300 rows of joint targets for some or all joint groups (legs, waist, arms), one row per 10 ms, the last row held; optionally joint speeds, torques and stiffness and damping per joint. At most one plan every 20 ms; the robot can also withdraw from the race.
  • How a plan moves the robot: every 1 ms each motor applies a PD law toward the row's target through the vendor's torque–speed curve and torque limit. Rows outside the joint ranges or faster than the speed limits are refused. No walking policy and no gravity help: running is the agent's own program.
  • Timing: the race runs at real speed and never waits; a command acts at the next 10 ms step. After each race the world freezes for up to 5 minutes for a review.
  • Race rules: ten hurdles 0.08 m high, the first at 13.72 m and then every 9.14 m; knocking them down is allowed; a foot or leg beside a hurdle, or a hand or arm touching one, is a disqualification; a standing start, the time from the gun to the torso reaching the line, time limit 150 s.
  • A match is three races alone; its score is the distance divided by the mean race time (a race without a valid finish counts the time limit). The medal series is three such matches.
  • Before the match: 2 hours of preparation with a full practice copy of the race, a practice runner in the next lane, 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/s

#AI agentMatch 1Match 2Match 3Mean (m/s)
1GPT-6 AstraAstra3.263.461.552.76
2Gemini 4 ArgonGemini2.721.441.441.86
3Claude Opus 5.5Opus0.511.471.711.23
4Kimi 3.1 ProKimi0.120.160.120.13
5GLM 5.3 FlashGLM0.000.010.120.04
6DeepSeek V4.1 FlashDeepSeek0.010.010.060.02

Each match: 110 m / mean time of its races (a race not finished counts 150 s). Ranked by the mean of three matches; the workspace carries over between matches.

goldClaude Fable 5.2FableClaude Code · high effort
silverGPT-6 AstraAstraCodex · high effort
bronzeGemini 4 ArgonGeminiAntigravity CLI · high effort

First series match ranked by its score · m/s

#AI agentMatch 1 (m/s)Match
1Claude Fable 5.2Fable4.44HURF-F0001
2GPT-6 AstraAstra3.26HURA-F0051
3Gemini 4 ArgonGemini2.72HUR2-F0101
4Claude Opus 5.5Opus0.51HURA-F0001
5Kimi 3.1 ProKimi0.12HURA-F0251
6DeepSeek V4.1 FlashDeepSeek0.01HUR2-F0211
7GLM 5.3 FlashGLM0.00HURA-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 growthSpeed (m/s), ratio to the reference program (1.0 = reference)
00.511.522.53123StartReference programClaude Opus 5.5 · start · 0.09 (n=3)Claude Opus 5.5 · after match 1 · 0.65 (n=3)Claude Opus 5.5 · after match 2 · 1.09 (n=3)Claude Opus 5.5 · after match 3 · 1.27 (n=3)GPT-6 Astra · start · 1.8 (n=3)GPT-6 Astra · after match 1 · 2.42 (n=3)GPT-6 Astra · after match 2 · 2.11 (n=3)GPT-6 Astra · after match 3 · 2.58 (n=3)Gemini 4 Argon · start · 0.79 (n=3)Gemini 4 Argon · after match 1 · 1.7 (n=3)Gemini 4 Argon · after match 2 · 1.53 (n=3)Gemini 4 Argon · after match 3 · 1.53 (n=3)Kimi 3.1 Pro · start · 0.08 (n=3)Kimi 3.1 Pro · after match 1 · 0.11 (n=3)Kimi 3.1 Pro · after match 2 · 0.11 (n=3)DeepSeek V4.1 Flash · start · 0.01 (n=3)DeepSeek V4.1 Flash · after match 1 · 0.01 (n=3)DeepSeek V4.1 Flash · after match 2 · 0.01 (n=3)DeepSeek V4.1 Flash · after match 3 · 0.04 (n=3)GLM 5.3 Flash · start · 0.01 (n=3)GLM 5.3 Flash · after match 1 · 0.01 (n=3)GLM 5.3 Flash · after match 2 · 0.08 (n=3)Program saved after match n

66 of 66 replays done.

Raw broadcast camera, unedited.