Asapi
For AI labs / Catalogue / Four‑car race driven from cockpit images on a live physics clock
Environment datasheet · Motor racing

Four‑car race driven from cockpit images on a live physics clock

ID apex/task01 · revision 1 · results as of 9 November 2026

Four agents share one race in a 3D physics simulation. Each sees only its own car's cockpit view and dashboard, and sends throttle, steering and brake while the world keeps moving; one valid lap earns the point.

Checks
1, a completed lap
Tested
4 models, one attempt each
Best
100% (4 models)
Median
100%
Full solves
4 of 4
Median attempt
12 min
What it tests

One lap from the driver's seat, on a clock that never pauses

Robot control

Each input holds for up to five seconds while the car moves on, so steering must allow for a picture already seconds old.

Visual spatial understanding

Each driver is given only its own cockpit image and dashboard to judge its place on the road and the next bend.

How it works

A cockpit image, three controls and a race that keeps moving

The agent's world

Chat tools to look at its own car's cockpit image and dashboard and send one input at a time; no map, route or coordinates.

Control

Throttle, steering and brake, held for up to five seconds; a fresh image is needed before the next input.

Getting stuck

A car stuck for eight seconds can ask to be reset; it keeps its progress, and 20 seconds are added to its time.

The reward

One point for a completed valid lap and none for a retirement; race time and resets are recorded but don't count.

Results

In one shared race, every attempt completed the lap; with reset penalties, race times ran from 6 to 36 minutes

Each dot is one attempt, and the line marks the median.

MedianFully solved
4 modelsOne attempt each100%4 of 4
ModelReward
GPT‑6 AstraOpenAI100%
GPT‑6 LunaOpenAI100%
GPT‑6.1 SolOpenAI100%
GPT‑6.1 TerraOpenAI100%
What models get wrong

One driver gets stuck 13 times, and only two never need a reset

Robot control

  1. By their own descriptions, every driver ran into a wall or barrier at least once and had to back away from it.

  2. GPT‑6 Luna got stuck 13 times, each reset adding 20 seconds, and took 36 minutes of race time, against 6 to 12 for the others.

  3. GPT‑6.1 Terra needed 2 resets on the way round; GPT‑6.1 Sol and GPT‑6 Astra needed none.

Running it

What an attempt takes, and what the environment includes

An attempt

Each car's race took 6 to 33 minutes of real time and from 87 to over 350 inputs.

What's in it

The race simulator, which now always models crash damage, the drivers' instructions, the completion check, and the race's recordings and replays.

Format

Runs on Node.js with a browser for the renderer; our races ran on a Windows PC, with other packaging on request.

Fresh variants

Other grids and races in the same design, made on request and never published.

The files and grader aren't published; the job's row on The AI Agent Trials, our public benchmark, sums up the job and shows each model's score.

Try one of our three sample environments

Three environments are packaged as samples you can download, with their graders and reference solutions. Ask for them with your work email, or book a call about the rest of the catalogue.