Screwed five‑board grid from angled stock by a bimanual robot
ID llarms/task05 · revision 2 · results as of 29 October 2026
In a simulated robot workshop, the agent drives a two‑handed robot by typed commands to unroll, tilt and turn five boards out of angled stock into a flat grid, each exactly inside its outline, and fix all six overlaps with three screws each.
- Checks
- 1, 1 point for the whole build, none for part of it
- Tested
- 1 model, one attempt
- Best
- 0% (GPT‑6.1 Sol)
- Full solves
- 0 of 1
- Attempt
- 20 min
Five boards laid as a flat grid, with six overlaps screwed
Robot control
Five boards come out of angled stock into a flat grid, and all six overlaps need three screws each: 18 screws, one at a time.
A simulated workshop, two hands, two drills and five boards in angled stock
The agent's world
A coding agent sends JSON to a local robot server: read the state, or send a motion of 1 to 32 velocity segments.
The robot
Two seven‑joint manipulators on sliding bases, each with a five‑fingered hand that holds by friction alone, so boards can slip, turn or drop.
Limits
One session of 20 minutes of wall time, planning included, with action and simulation‑time budgets; no resets or direct state changes.
The reward
One point only if all five boards settle inside outlines of their own size and all six joints are rigid with three screws each.
All five boards ended outside their outlines, one by 16 cm, and no screw went in
The dot is the one attempt.
| Model | Reward |
|---|---|
| GPT‑6.1 SolOpenAI | 0% |
This attempt was told an earlier, looser target that accepted boards centimetres off, and not the corner rule; its score uses the strict outlines. A person can complete this build easily.
It leaves the grid short, one board 16 cm out, and drives no screw
Robot control
All five boards came to rest outside their outlines, by 24.1, 62.2, 27.1, 30.3 and 161.1 mm at the worst corner.
Placement never led to a full fastening cycle, so none of the 18 screws went in and none of the six joints was rigid.
The server stopped the attempt at the 20‑minute limit, after 55 motion commands covering 343 simulated seconds.
What an attempt takes, and what the environment includes
An attempt
Cost and tokens weren't recorded: the attempt ran on a subscription sign‑in, which logs neither.
What's in it
The frozen runtime with this scenario, the request and interface document, the grader and its self‑test, and the attempt's full record and videos.
Format
Node.js, with the physics in headless Edge and a Python grader; it runs on a Windows PC today, and other packaging is on request.
Fresh variants
Other builds and layouts 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.
AsapiThe AI Agent Trials