Reinforcement learning you can see and touch

From simulated training to a physical track, every stage of learning is observable, operable, and testable.

Reinforcement learning is a loop

You do not write the model. You run it, look at what happened, change something, and run it again.

  1. Design

    Decide what the car is allowed to do, and what counts as doing it well. The action space and the reward function are the only instructions you can give a model, which makes this where the loop really begins.

  2. Train

    Try, fail, and try again in the 3D simulator, with PPO learning from each attempt. Crashing costs nothing, and it all runs on your own machine.

  3. Deploy and drive it

    Send the trained model to a TopRacer Car and let it drive TopRacer Map 2018 for real. Nothing has to be rewritten to make the trip.

  4. Evaluate and think

    Read the lap times, the driving path, and the mistakes to work out what actually happened, then decide what to change. That thinking is worth more than another round of training.

Back to step one The new settings go back to step one and train another round. The loop runs as long as you want it to, and every round you can see how much better it got.

The product family

The software trains and evaluates. The hardware takes the result out into the real world and tests it.

Mobile application Free

MiniDeepCar

Set up the car over Wi-Fi or Bluetooth, deploy a model, check vehicle status and the live camera feed, or drive it yourself.

Physical vehicle

TopRacer Car

Takes the model you trained in Console out onto a real track.

USD 699 each

Coming soon

Physical track

TopRacer Map 2018

A physical track to use with track analysis in TopRacer Console.

USD 599 per set

Coming soon

Service

Online race hosting

We open an online race for you, and entrants upload their runs straight from TopRacer Console.

USD 20 per race

Why seeing it matters

Reinforcement learning is hard to learn because what happens in the middle is invisible. TopRacer opens up every stage.

Every step leaves a record

Training curves, simulator replays, and the path the real car drove -- all in the same tool.

Simulation and car run the same model

What you trained in the simulator is what goes on the car. There is no second system in between.

It runs on your own computer

Training and evaluation are local. No cloud account, no rented compute.

Schools and groups

TopRacer Car and TopRacer Map 2018 are coming soon. If you are evaluating them for a course or a club, tell us what you need.