FrostPath turns live sport into real-time, predictive graphics — grounded in full-fidelity physics and driven by adaptive deep reinforcement learning. It reads the game from the camera feeds you already run, forecasts what happens next, and renders it broadcast-ready. One platform: cloud-trained, edge-deployed, and scalable to any venue.
Request a Demo See it on the broadcast →Legacy shot-tracking is hardware-intensive — dedicated sensors, on-site rigs, costly installs, and crews on the ground before a single frame airs. FrostPath removes all of it. The only inputs we need are the overhead and broadcast camera feeds already covering the sheet.
Because the intelligence lives in the model — not in on-venue equipment — FrostPath deploys remotely. No sensor array to mount, no calibration truck to dispatch, no camera crew to fly in. Point us at the existing streams and the platform does the rest, so we can light up a new arena or event without ever setting foot in it.
We ingest the overhead and broadcast feeds already in production. Nothing gets added to the rig.
No LiDAR, no tracking beacons, no embedded chips. Computer vision reads the play directly from video.
No on-site build, no calibration day, no venue downtime. Provisioning is remote and software-defined.
Deploy to any venue from the cloud, and scale to the next event without shipping a kit or a team.
Real broadcast footage of rare, decisive moments is scarce, unlabeled, and slow to collect venue by venue. FrostPath sidesteps that bottleneck: our models learn from synthetic, physics-simulated data generated by a full-size engine — so training coverage is a choice, not a limitation.
At the core is a full-size 40 m curling physics model computing 24,000 stone-and-pebble interactions per second, built on the most widely accepted friction-physics research. It produces every trajectory, ice condition, and edge case the game can throw — including situations a camera might see once a season — as clean, perfectly labeled training data.
Simulation generates the outlier shots and marginal geometries that almost never appear in captured footage, so the model is ready for them on air.
Training data is generated on demand, not waited on. New scenarios and refinements ship in the loop instead of over seasons.
Because the physics is universal, models need no bespoke footage from every arena. Deploy anywhere without a data-gathering campaign.
Physical AI means intelligence grounded in the real world — models that don't just pattern-match pixels but reason about forces, trajectories, and dynamic conditions, then learn optimal responses as those conditions change. That grounding is what makes real-time prediction, not just replay, possible.
Rules-based analytics and static ML can describe what already happened; they can't forecast a sport that rewrites its own conditions shot to shot. In curling, the ice changes every game, every end, every rock — and a shot's true value is only knowable once the end is scored. FrostPath's adaptive agents are built for exactly that kind of moving target.
Deep reinforcement-learning agents learn optimal responses to dynamic, changing conditions in real time. Two curling agents are trained, with an adaptive framework ready for broadcast integration.
The 40 m engine and its 24,000 interactions/sec give every prediction a physically accurate foundation, from shot trace to end forecast to expected value.
Live and simulated model data is rendered into multi-user, full-fidelity 3D — the bridge from model output to broadcast-ready graphics, built on NVIDIA Omniverse.
FrostPath is engineered on the platforms trusted to train, secure, and render at scale — so the pipeline from cloud training to on-air graphics is robust, repeatable, and production-ready.
The training and operations backbone: RL/ML training pipelines, an MLOps model registry with CI/CD, and a secure artifact store on S3 and ECR. Cloud-to-edge delivery pushes model updates without taking a broadcast down.
GPU acceleration from training to inference. FrostPath is a member of the NVIDIA Inception program, with access to the partner ecosystem and hardware built for real-time, high-fidelity AI.
The full-fidelity visualization and collaboration layer, built on OpenUSD. It turns live and simulated model data into immersive, broadcast-ready 3D — the same environment used to build our 4K arena VFX.
FrostPath runs a centralized-production, remote-delivery model: heavy compute stays in the cloud, security is enforced at the boundary, and lightweight inference runs at the venue on the existing feeds. It's the model proven at scale across modern cloud-based sports-graphics delivery — one production pipeline, many venues and events, no crew on the ground.
Where synthetic data becomes deployable intelligence. Models are trained and versioned on the AWS stack, then packaged for delivery to the edge.
Every edge-bound model passes a validation and security gateway — compliance and monitoring checks, air-gapped staging, and signed-boot model verification — before it's cleared for a venue.
Inference runs at the venue on the existing camera streams, with a local control loop and digital-twin integration driving the live graphics. Cloud-to-edge MLOps updates models without downtime.
Because every stage above the venue is centralized and software-defined, adding an arena is a deployment, not a build-out. The architecture scales across events on the streams already in place.
No new hardware. No install day. No crew on the ground. Bring adaptive Physical AI to your venue and see it live.