NVIDIA Cosmos Cookoff 2026

AI-powered coach, referee & guide

Professional-grade officiating and real-time coaching for every level of play — from training sessions to tournaments, powered by physical AI reasoning.

5 Sports
Edge On-Device AI
CoT Physical Reasoning
Real-Time AI Analysis
See how it works

From Camera to Scoreboard

Point your phone. The AI watches the game, reasons about ball physics, and updates the score — all in under 2 seconds, entirely on device.

Camera Feed

Phone or IP camera captures the game via OpenCV feed

Frames

Jetson AGX Orin

NVIDIA edge GPU runs Cosmos Reason 2 vision analysis pipeline

Keyframes

Cosmos Reason 2

Chain-of-thought physical reasoning: trajectory, spin, boundary detection

Events
42

Mobile App

React Native app shows live scores, AI reasoning, and commentary

Protocol WebSocket + REST
Model nvidia/cosmos-reason2-8b
Format OpenAI-compatible API
Deployment HuggingFace + LiteLLM

Five Sports. One AI.

Tennis, pickleball, badminton, table tennis, and cricket — each with sport-specific physical reasoning tuned for the rules and physics of the game.

🎾

Tennis

In/out calls, shot type detection, serve placement, spin & rally analysis

🏓

Pickleball

Kitchen violations, shot classification, serve legality, double-bounce rules

🏸

Badminton

Shuttlecock trajectory, smash detection, boundary calls, serve compliance

🏓

Table Tennis

Spin analysis, edge detection, serve legality, let calls, rally patterns

🏏

Cricket

Trajectory, boundaries, wickets, shot classification, swing & spin analysis

Watch It Think

Cosmos Reason 2 analyzes each frame with step-by-step physical reasoning — trajectory, spin, boundary detection — then delivers a structured decision with confidence scores.

cosmos_reason2_output.json nvidia/cosmos-reason2-8b
<think>
Looking at this cricket frame, I need to analyze the ball trajectory and determine the scoring event. 1. The ball has been bowled and the batsman has played a shot. The bat angle suggests a cover drive - face open, leaning into the stroke with front foot forward. 2. Ball trajectory analysis: The ball appears to be traveling along the ground toward the off-side boundary. Speed estimation from motion blur suggests ~80 km/h post-contact. 3. Fielder positions: I can see a cover fielder diving but the ball has passed to their right. No other fielder in the path between ball and boundary rope. 4. Physics check: At this speed and angle, the ball will reach the boundary rope within ~2.5 seconds. No obstacle in path. Probability of boundary: HIGH. 5. Delivery analysis: The bowler delivered a length ball on off-stump. Slight outswing visible from seam position. The batsman has used the pace of the ball effectively.
</think>
<answer>
{
  "phase": "batting",
  "ball_state": "traveling_to_boundary",
  "event": "boundary_4",
  "confidence": 0.92,
  "description": "Cover drive through the gap, ball racing to the boundary. Clean stroke, no fielder in the path.",
  "shot_type": "cover_drive",
  "delivery_type": "outswing",
  "speed_estimate": "medium"
}
</answer>
1

See

Camera frame is sent to Cosmos Reason 2 with sport-specific prompts

2

Think

Model reasons about ball physics, trajectory, player positions step by step

3

Decide

Structured JSON output with event classification and confidence score

4

Score

Event feeds into RefereAI and broadcasts live scores to connected devices

Live Reasoning

Pick a sports clip below and watch Cosmos Reason 2 analyze it in real time. See the chain-of-thought reasoning as the AI figures out what happened.

Selected frame

Select a frame and click Analyze to see Cosmos Reason 2 in action

Your AI Companion

Real-time scoring, AI commentary, and physical reasoning — delivered entirely on-device with zero cloud dependency.

Physical Reasoning

Cosmos Reason 2 analyzes ball trajectories, bounce physics, and player positions with chain-of-thought reasoning. The model thinks step-by-step about what it sees before making a scoring decision.

Cosmos Reason 2 Chain-of-Thought Edge AI

AI Commentary

Live play-by-play commentary generated from scoring events. Context-aware descriptions of shots, deliveries, and rallies. Multiple commentary styles for casual and competitive play.

English Context-Aware Real-Time
42

Live Scoring

Automatic score updates via WebSocket. Cast to any screen as a TV-style scoreboard. Works offline with local SQLite storage.

Multi-Sport Engine

Sport-specific reasoning for cricket, tennis, pickleball, badminton, and table tennis. Each with tuned Cosmos physical AI prompts.

Powered by Jetson AGX Orin

275 TOPS of AI compute in a portable module. Deploy at training facilities, clubs, or tournaments — full inference on-device, no cloud required.

ORIN
NVIDIA Jetson AGX Orin 64GB
GPU 2048 CUDA cores + 64 Tensor cores
Memory 64 GB LPDDR5 (204.8 GB/s)
AI Perf 275 TOPS (INT8)
CPU 12-core Arm Cortex-A78AE
DLA 2x NVDLA v2.0
Power 15W - 60W (configurable)
Serving HF Transformers + LiteLLM
Pipeline FastAPI + Tailscale Funnel

Why Edge AI?

Low Latency

On-device inference, no cloud round-trip needed.

Works Offline

No internet needed. Play anywhere -- parks, backyards, remote fields.

Privacy First

Video never leaves the device. All processing is on-device.

Cost Effective

No per-inference cloud costs. One-time hardware investment.

The Team

Built by engineers and strategists who believe every player deserves professional-grade tools.

Ravinder Jilkapally

Ravinder Jilkapally

Product & Engineering Lead

Scaling AI at Xpanse · Startup things at aisoft.us

Product, platform, and engineering leader with 18+ years building large-scale AI/ML platforms, GenAI products, and cloud-native data systems. Building the future of sports officiating with physical AI — your personal coach, referee, and guide, powered by NVIDIA Cosmos.

Harsha Kalapala

Harsha Kalapala

Growth & Strategy

Sr. Director, Segment Marketing @ AlertMedia

Experienced marketing leader who makes complex things simple. Product marketing, go-to-market strategy, and building communities around technology products.

Built for the NVIDIA Cosmos Cookoff — a hackathon challenging developers to build physical AI applications with Cosmos Reason 2.

You've used your free tries

0 / 5 live inferences remaining

Each live request runs Cosmos Reason 2 on our Jetson AGX Orin in real time. Star the repo and follow for updates on unlimited access.

Star on GitHub