Two boards placed together by a bimanual robot, one per hand

Two boards, one per hand, both exactly inside their outlines at once

Role

The controller of a simulated bimanual robot in a workshop of wooden boards, acting only through a local HTTP command service, with nobody giving movement advice.

Request

Pick up both boards, turning them if needed, and leave each settled inside its outline with every corner in, meeting the lift, release and settling conditions; check the measured result before stopping. One engine start, no reset.

Deliverable

The final physical state, from the engine's last observation: both boards lifted by their own hands, released and settled. A short summary is saved too; its claims don't count.

What the reward depends on

All eight corners of both boards inside outlines of their own size, settled at the same time, each after a real lift by its assigned hand. One point or none: coming close earns nothing.

Where the requirements are

Stated in the request and one interface document, with the runtime's source open to read; each observation reports the targets and progress. Nothing is hidden: the difficulty is physical control.

A contact‑only physics workshop, read as exact state and driven by typed commands

Interface

A Codex chat subagent running GPT‑6.1 Sol at its default reasoning setting, sending JSON to a local HTTP control server through its shell tools: read the state, send a step or a sequence of velocity segments, and post short progress notes.

Physics

Rigid‑body physics (cannon‑es) in its own headless Edge process at 180 Hz, apart from any viewer. Simulated time advances only during commands and pauses while the agent plans.

The robot

Two seven‑joint manipulators on stationary bases, each with a palm and five fingers of three joints, with spread and thumb opposition. Motion is rate‑limited and stopped by collisions, so a target isn't always reached.

Grip

No attachment constraint: a hold is opposed, force‑bearing finger contact, so a board can slip, turn or drop while the hand stays closed, and over‑curling can push it out.

Observations

Exact structured state, not images: board poses, sizes and velocities; hand and finger poses, contact force and slip; collision and joint‑limit feedback; each target's pose and progress.

The build

Two boards on the deck, one assigned to each hand, each to be lifted by that hand and released in its matching outline; both must be placed together. No tools or screws.

Recording

Every command result and sampled pose is saved; four fixed cameras render review videos afterwards, at simulation speed. The saved observations, not the videos, are the record.

Network

Internet on, for the pinned physics libraries, and web tools available; nothing the job needs is online.

Limits

One session of 20 minutes of wall time, planning included, plus the scenario's own action and simulation‑time budgets. The request rules out resets, teleporting, direct state changes and earlier trajectories.

One all‑or‑nothing check on the final physical state

Structure

One check, worth one point: the whole build complete. Boards placed and each board's worst corner protrusion are reported as diagnostics, never scored.

Placement

Each board must be lifted, released and left to settle with all eight corners inside a target outline of its own size, allowing only a numerical margin. Holding a board in place doesn't count.

Who decides

A script reads the final observation and transforms every board's eight corners into its target's frame; it also needs the engine's own success flag for lift, release, settling and any joints. No model judge, and the agent's claims don't count.

The rule in the request

The request stated that every corner must lie inside its outline, with the margin for rounding only, so this attempt was told the strict rule.

Engine reward

After each action the engine also returns a shaped reward from the boards' distances to their targets, with a bonus on success and a penalty for a lost board; the grader ignores it.

Formula

Reward = 1 if the build is complete, else 0. An untouched scene, a partial build or a missing final state scores 0.

Both boards ended millimetres out when simulation time ran out

GPT‑6.1 Sol, the only model tested on the builds, scored 0%. It carried both boards to their outlines, but when the simulation‑time budget ran out, after 12 minutes, they rested 4.2 and 9.5 mm out at the worst corner.

Each dot is one attempt; the line marks the median.

One attempt, by GPT‑6.1 Sol at seed 1; no other model has been tested on the ten builds.

Made on 2 November 2026 on its own control server and physics process, told the strict rule in the request. Its closing summary called the build incomplete, as the grader found.

What the model got wrong

One attempt, by GPT‑6.1 Sol, told the strict rule; command numbers are from its saved sequence.

1. Millimetres out. Both boards came to rest outside their outlines, by 4.2 and 9.5 mm at the worst corner; no saved state had either fully inside, the closest 1.7 mm out.

2. Corrections that didn't converge. Commands 13 to 31 alternated finger moves, wrist turns and pushes, many blocked by collisions; releases and contact kept disturbing the alignment.

3. A second placement, still out. Commands 32 to 38 regrasped and released both boards again, and the last two used up the simulation‑time budget, with 8 of the 20 minutes left.

What an attempt costs

Length

The attempt took 12 minutes of wall time and 40 motion commands, using its whole simulation‑time budget. Model calls, tool calls and tokens weren't recorded; a motion command isn't a model call.

Cost

Not recorded: the attempt ran on a subscription sign‑in, which logs no cost or tokens, so we give no figure.

What you license

With a licence

Revision 1: the frozen runtime with this scenario at seed 1, the request and interface document, the grader and its self‑test, and the attempt's commands, observations, pose trajectory and four review videos, as baseline results.

Format

Runs on Node.js, with the physics in headless Microsoft Edge driven by Playwright, and a Python grader; today on a Windows PC, outside a container. Other packaging on request.

Fresh variants

Other builds and layouts in the same design, on request. They're never published.

Public or not

The AI Agent Trials, our public benchmark, sums up the job and shows each model's score, so hold out fresh variants for evaluation. The files, grader and answer key aren't published.

What we haven't measured, and what's uneven

• One attempt, by one model, at one seed, so we haven't measured how much attempts vary or how other models do.
• No reference solution: no attempt, scripted or by an agent, has completed a build under the strict check, so it isn't shown that one can be; the self‑tests check only the scoring.
• Each outline is its board's own size, so a board must settle almost exactly at its target pose; whether the contact solver allows that hasn't been tested.
• The grader trusts the engine's own flag for lift, release, settling and joints and checks only the geometry; regraded copies with that flag forced true still failed.
• The agent reads exact state and may read the runtime's source, which reveals the thresholds: this tests control from state, not from vision.
• Simplified physics: rigid contacts and simplified fingers, so results don't speak to real robots.
• Up to three attempts shared one PC at a time, so wall time may include contention.