Asapi
For AI labs / Catalogue / Four‑car race on a live physics clock with lasting crash damage
Environment datasheet · Motor racing

Four‑car race on a live physics clock with lasting crash damage

ID apex/task02 · revision 1 · results as of 10 November 2026

Four agents share one race in a 3D physics simulation, each driving from its own cockpit view. Crashes permanently cut power, braking, grip and steering, a reset never repairs them, and critical damage retires the car; one valid lap earns the point.

Checks
1, a completed lap
Tested
4 models, one attempt each
Best
100% (3 models)
Median
100%
Full solves
3 of 4
Median attempt
18 min
What it tests

One lap from the driver's seat, where every crash costs for good

Robot control

Crashes weaken the car for the rest of the race, so inputs must keep it off the walls as power, grip and steering fall.

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, a damage gauge and a race that keeps moving

The agent's world

Chat tools to look at its own car's cockpit image and dashboard, now with a damage gauge, and send one input at a time.

Damage

Crashes permanently reduce engine power, braking, grip and steering; a reset never repairs them, and critical damage retires the car.

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, resets and damage are recorded but don't count.

Results

In one shared race, three attempts completed the lap, and the fourth retired when its engine failed

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

MedianFully solved
4 modelsOne attempt each100%3 of 4
ModelReward
GPT‑6 AstraOpenAI100%
GPT‑6.1 SolOpenAI100%
GPT‑6.1 TerraOpenAI100%
GPT‑6 LunaOpenAI0%
What models get wrong

One crashes until its engine fails, and even the cleanest drives take knocks

Robot control

  1. GPT‑6 Luna crashed 21 times and needed 18 resets; its damage reached 100% and the engine failed before it completed the lap.

  2. GPT‑6.1 Terra crashed 12 times and needed 13 resets, finishing with 34% damage and a race time, with penalties, of 27½ minutes, against 10 for GPT‑6.1 Sol.

  3. Even the two cleanest drives, GPT‑6.1 Sol's and GPT‑6 Astra's, crashed 3 and 5 times.

Running it

What an attempt takes, and what the environment includes

An attempt

Each car's race took 10 to 23 minutes of real time and 56 to 80 inputs.

What's in it

The race simulator with its renderer and physics, 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 damage settings 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.