precision_manufacturing Sideline Launch Referee arrow_forward
// MISSION_ORBIT_INITIATED

An AI referee.
Watch it make
the call.

240 million tennis matches per year. Almost none have a referee. Sideline watches video, reasons through the play with chain-of-thought, and signals a physical robot.

240M
Matches / yr
72B
VLM params
3
Embodiments
AXIS 01 / SCALE 1:1
REF-SIM_BETA
NEBIUS.BUILD_SF · 2026
OUT →
SO-101 · OBSERVING
SO-101 Arm — Ready
240M matches / yr / 72B VLM params (Qwen2-VL) / 3 physical embodiments / Apache-2.0 Open Source
// SYSTEM_ARCHITECTURE

Camera to call, in one loop.

An image classifier can tell you the ball landed out. An agent knows the score, the server, the rules, and what happened three points ago.

photo_camera
01 / INPUT

Camera

Multi-stream ingest, 30fps sync, H.264/265.

neurology
02 / VLM

Nebius VLM

Qwen2-VL-72B via Nebius Token Factory.

psychology
03 / AGENT

CoT Agent

Chain-of-thought + match-state memory.

terminal
04 / DISPATCH

Tool Call

Serialized commands to robot runtime.

precision_manufacturing
05 / ACTION

Gesture

Physical point + audio annunciation.

// WHY_NOT_A_CLASSIFIER

Classifier ≠ referee.

A classifier tells you the ball is out. A referee holds state, narrates reasoning, and takes physical action.

// CORE_ADVANTAGE_01

Chain-of-thought

Every decision is narrated. "Ball crossed the plane at 47 km/h; angle -12°; topspin — goal." The agent leaves a trace a human can audit.

DEBUG_LOG_SAMPLE
>> ball_velocity = 47.2 km/h
>> trajectory_angle = -12.4°
>> rule_check = within_bounds ✓
>> DECISION: GOAL (CONF: 0.94)
// CORE_ADVANTAGE_02

State memory

Score, server, faults, who won the last point — all in context. Every call incorporates match history, not just the current frame.

// CORE_ADVANTAGE_03

Multi-VLM swap

Hot-swap between Nebius Qwen2-VL-72B, Cosmos Reason, or local Llama without touching agent logic.

NEBIUS_QWEN2VL COSMOS_REASON LLAMA_LOCAL
// CORE_ADVANTAGE_04

Physical grounding

Direct tool-call to robot runtime. The referee doesn't just show a score — it points, speaks, and gestures.

// PHYSICAL_EMBODIMENTS

Three robots. One agent.

The referee reasoning stays the same. The embodiment swaps: precise arm, humanoid presence, mobile rover.

precision_manufacturing
EMBODIMENT 01

SO-101 Arm

READY
DOF6-AXIS
SDKLEROBOT · ROS2
ROLEPOINT · RAISE · GESTURE

Six-joint manipulator. Precise court-side gestures, high-signal low-footprint. Perfect for in-stadium calls.

accessibility_new
EMBODIMENT 02

Unitree G1

SHARED FLEET
TYPEHUMANOID · 23 DoF
SDKUNITREE_API
ROLESTADIUM GESTURES

Full-humanoid gestures — arm raise, point, whistle posture. Stadium-scale presence when the call needs to be seen.

smart_toy
EMBODIMENT 03

MentorPi

MOBILE
PLATFORMHIWONDER ROVER
SDKROS2 · PYTHON
ROLEFOLLOW + ANNOUNCE

Tracked rover with mecanum wheels, camera, speaker, and ROS2. Mobile, affordable, follows the play.

// BLUEPRINT_SPEC_SHEET

Engineering readout.

// REV_A · 2026-04

SHEET 01 / 01
Component Specification Status
VLM Backend Nebius Qwen2-VL-72B · OpenAI-compatible endpoint Active
Agent Runtime Python 3.12 · CoT prompt · tool-calling · match-state memory Active
Rules Engine Tennis (MVP) · pluggable sport config Active
Robot IO ROS2 · LeRobot · gesture scripts · tool-call → execute Testing
Console Web console · live token stream · CoT trace · tool-call log Active
License Apache-2.0 · fork it · build your own sport Open
// BUILT_AT_NEBIUS.BUILD_SF_2026

Four engineers, one weekend.

Watch it call the point.
Fork it. Build your sport.