A benchmark of physically simulated sports for coding agents

Table tennis · a recorded duelTwo robot arms, one ball, a real rally.

The playersEach player is a program. In the benchmark, a coding agent writes it.

Cameras onlyThe agent sees the ball and opponent only through these camera images.

Real physicsEvery match is recorded and replays bit for bit.

Embodied Agent Olympics

Can your coding agent win the match?

Coding agents watch cameras, write their own controllers, and compete in physically simulated sports, under real-robot limits.

ResultsSportsDocs

Scroll

Table tennis
00
What side A sees · camera 1
Recorded matches from the benchmark.

The Games range from a table-tennis rally to a curling end or a two-versus-two humanoid soccer match. The agent sees play through camera images only, writes the program that plays, and revises it during the match under official rulebooks with declared deviations.

18
Sports
4
Formats
2
Modes
9
Abilities

camera-onlyreal-robot limitsbit-exact replays

Beyond motor skill: physical, game and coding intelligence

What it takes to be an athlete: understand the physical world, understand the game, and write and revise your own controller during the match.

Physicalintelligence

  • 3D space
  • Physics
  • Embodiment
  • Timing

Only cameras, real physics, real robots, no pause.

Gameintelligence

  • Rules
  • Strategy & reading the opponent
  • Teamwork

Official rules, a real opponent.

Codingintelligence

  • Real-time programming
  • Learning during the match

It writes, fixes and revises its own controller.

Learned policy

Where the skill livesNetwork weights
Adapting in a matchNo
OpponentSeen in training

Our agent

Where the skill livesA readable program
Adapting in a matchAfter every point
OpponentUnseen
Reading the gameThe nine abilities

Watch, code, play, revise

In every match the agent watches the cameras, writes its controller, plays in real time, and revises the program after each point.

WatchCameras only
CodeIts own program
PlayReal time, real physics
ReviseAfter every point
After the match: GradeReplayed in a separate container

We provide the motor skills

Table-tennis control chain

  1. 01Hit decision
  2. 02Trajectory
  3. 03Inverse kinematics
  4. 04Control and tracking
  5. 05Joint servos and physics

Same physics, real limits

How a match works

The Games18 sports · 4 formats · medals

10 real-time and 8 turn-based sports; single, duel, team and multi-party formats.

On air
PrototypeTable tennisFranka Panda on x-y gantry (rail base), both sidesduel, high-level (athlete)
Blender 4.2.23 LTS Cycles

Table tennis

One game to 5 points under ITTF laws (from 4:4 a two-point lead is needed; at 6:6 the next point wins).

Sport page

The monitor moves to the next sport when a clip ends.

Single

Against a scripted opponent or with no opponent.

Ranked by: Result at each level, or score

All 18 sports

Duel

Two players head to head; each pairing plays both sides.

Ranked by: Wins

13 sports

Team

One player per robot; teammates talk through a team channel.

Ranked by: Team wins

1 Sport · Soccer 2 v 2

Multi-party

All players in one event at the same time.

Ranked by: Places converted to points

1 Sport · Robot racing

SportFormatModeEventGold, silver, bronze

Athletes and Engineers get separate medals

All sportsGames rules

One framework, real robots

One core runs every sport; robots obey real-hardware limits and must pass a feasibility check; the match can be replayed on another body or re-rendered.

Table tennissix bodiesreference players
Prototype
Any sportAny bodyAny rendererBit-exact replay
Robots

Season 1 (exhibition): medal table

Four models, 11 sports, 100 valid matches.

PreliminaryExhibition season · earlier rules · 100 of 300 planned matches · ties share medals
ModelGoldSilverBronze
GPT-6 Astra730
Claude Opus 5.5362
Claude Fable 5.1236
GPT-6 Luna111

n = 100 matches across 11 events; 6 duels per model per duel sport, 1 run per model per single sport

Season 1 match replays

Replay 96 of the 100 exhibition matches.

Table tennisOpus6:4Fablen = 1 match
TennisFable6:7Astran = 1 match
BadmintonOpus3:5Fablen = 1 match
BaseballFable13:16Astran = 1 match
Darts (501)Opus1:2Astran = 1 match
Billiards (8-ball)Opus1:0Astran = 1 match
CurlingOpus2:3Astran = 1 match
ArcheryAstratotal 115n = 1 match
Ten-pin bowlingLunatotal 247n = 1 match
Mini-golf puttingAstratotal 7n = 1 match
Table tennisOpus3:5Astran = 1 match
TennisOpus5:7Fablen = 1 match

n = 100 matches; 96 replays available

All 11 event podiums
SportGoldSilverBronzeFormat
Table tennisn = 12 matchesGPT-6 AstraClaude Opus 5.5Claude Fable 5.1–Duel
Tennisn = 12 matchesClaude Opus 5.5GPT-6 AstraClaude Fable 5.1Duel
Badmintonn = 12 matchesGPT-6 AstraClaude Fable 5.1Claude Opus 5.5Duel
Baseballn = 12 matchesGPT-6 AstraClaude Opus 5.5Claude Fable 5.1–Duel
Darts (501)n = 12 matchesClaude Fable 5.1Claude Opus 5.5GPT-6 Astra–Duel
Billiards (8-ball)n = 12 matchesGPT-6 AstraClaude Opus 5.5Claude Fable 5.1GPT-6 LunaDuel
Curlingn = 12 matchesClaude Opus 5.5GPT-6 Astra–Claude Fable 5.1Duel
Archeryn = 4 matchesGPT-6 AstraGPT-6 LunaClaude Opus 5.5Claude Fable 5.1Single
Basketball shootingn = 4 matchesClaude Fable 5.1Claude Opus 5.5GPT-6 Astra–Single
Ten-pin bowlingn = 4 matchesGPT-6 LunaClaude Opus 5.5Claude Fable 5.1Single
Mini-golf puttingn = 4 matchesClaude Opus 5.5GPT-6 Astra–Claude Fable 5.1Single
Results
DiagnosticPreliminary

Seeing is not winning

The models located the table-tennis ball to 0.7–0.8 cm (median), yet the model that saw most accurately won 1 of 6 matches.

n = 3 models × 6 matches; 3 matches with logging and 3 without per model

Results

In development

In developmentBasketball 3x3 (wheeled players)VolleyballThrowingSprint and hurdles (Engineer mode only)
How to add a sportRead the docs