Release notes¶

Entries are added per tag. The cloud app and the car have separate version lines.

Cloud¶

New

Your run plays back in 3D. When a run finishes, the trained policy drives in your browser: six cameras, a HUD with speed and the steering and throttle it chose, a minimap, scrub and speed controls. It plays in the lab panel, on your run page and on the Gym cards. Switch between the real car's shape and the plain box-and-four-wheels the simulator trains on.
Race. The new Race page puts you in the simulated car on the real track layouts: arrow keys or WASD, or a gamepad. Race a ghost of a posted lap, or the trained keeper policy driving live in your browser at the same 15 Hz it drives the car. A hit stops you at the wall; drive on. Three laps post your time, the server replays your inputs to validate it, and the best laps go on your team's board. Race your own ghost next time.
Your own trained policy can be the opponent: any finished Train for the Real Car run is in the opponent list.

Changed

No more rendered video clip. The playback is the real trajectory of the rollout, which is smaller, faster, and never black. Runs record it in under a minute on any GPU host.
Cloud runs land on the cheapest capable GPU first.

Fixed

The steering in the race matches the simulator: the plant was measured against the real training environment, and the default steering authority is the real car's teleop setting.

Fixed

Black clips. Your run now lands only on GPU hosts with a driver the renderer accepts, and a host that still cannot render is replaced before training starts, in another data center. The policy was never affected; the clip is.
Launches no longer fail when the first two GPU types are sold out: a third type is tried.

New

The Labs page shows the three training labs: First Training, More Training Is Not Better, Train for the Real Car. The vision and control labs come back later.
First Training and More Training Is Not Better train the task GoatRacer-Fiesta-3obs-v1: the policy sees three numbers, no padding. The network line and the lab copy agree. Old runs keep their policy file.
Flag this run: report a bad clip, a stuck run, a wrong SIL verdict or a broken page from the run panel, the run page or the lab header. The team reads every report.
A GPU host that cannot render is caught at boot and you are moved to another one before the run starts.
The training lab pages are one screen each.

Changed

Your run is charged from the moment the GPU has the job, not from the click. A slow host start is on us.
A warm GPU waits three minutes for your next launch. "Keep it warm?" adds three minutes from your balance.

New

The app moves you to another GPU host when one does not start. A host that has not answered in five minutes is ended and your run or warm-up is made again elsewhere, up to three times. The panel says so: "that host did not start - moving you to another one (attempt N)".
The warming panel says what it knows: machine assigned and pulling, image up and agent starting, moved.
Runs keep their error reason on the run row.

Fixed

A run queued on a booting GPU could be shut down by a second reader before it started. One student's run died that way on 2026-09-21.
A failed-start report ran twice per run.

TODO (Rob): add the current car runtime version and its customer-facing notes.