Skip to content
Effector Lab

Try every tool in physics before you buy one.

Describe the job your robot has to do. Effector Lab plans it as steps, tests every end effector it can simulate on every step, and recommends one tool, a tool change or a two-tool head, next to the best single off-the-shelf tool. This site replays recorded plans and runs, and can plan and test a new job live.

This site replays recorded plans, and you can build yours by hand. To plan new words, open the Studio and use your own key or a small free tier, or run it yourself.

Recorded physics run, not an animation

Step 1 of 4: Open the lid

Industrial servo parallel gripper, 110 mm stroke. Hook the lid's lip and pull it open past vertical.

What it shows: the recommendation for the returns tote, a single industrial servo parallel gripper, 110 mm stroke, on its recorded run (nominal placement, seed 0). 1. Open the lid; 2. Pick the carton; 3. Pick the pack; 4. Close the lid. Done in 78.2 s.
  1. 1Plan

    Describe the job in plain words, and the planner turns it into objects and steps the simulator can run. Here you replay recorded plans and edit their steps, or plan a new job live with your own NVIDIA or Token Factory key, or a small free tier.

    Returns tote: one paragraph, 4 steps, 4 objects

    Plan a task
  2. 2Test

    Every tool it can simulate tries every step it could do, in physics, on the same placements and seeds. Tools that cannot do a step are ruled out with a reason.

    56 tools, 249 screen runs, 10 minutes

    See a screen
  3. 3Recommend

    Single tools, tool changes and two-tool heads are combined from what passed, and the finalists run the whole task against the best single off-the-shelf tool.

    8 of 8 runs; the pick is the best single tool

    See a recommendation
  4. 4Build

    Download a build pack: the printed bracket or changer plates, a bill of materials, the tested scenes and a certificate.

    7 files, 11 kB

    See a build pack

Tested before we said it

Two pre-declared tests, frozen before their final runs, with the data in the repository.

Studio benchmark v1 (top-down wrist, 19 Oct): held-out conditions, three use cases

The Studio's pick74 of 96

Best single off-the-shelf tool0 of 96

Runs that finished the whole task, on conditions neither setup was chosen on, with the v1 setups. Returns tote: 32 of 32; Pipe coupling: 32 of 32; Equipment cabinet: 10 of 32.

v2 adds wrist tilt and honest tolerances; its benchmark is pending.

Read the benchmark

Finger designer, pre-registered

9fewer designs to a finger that seats the part

The language-model designer with feedback against equal-budget random search: 8.97 fewer on average (95% interval 3.5 to 14.2), exact permutation test p = 0.0011.

Read the result

Simulated results. They rank tools under stated conditions; verify on hardware before you rely on them.

An open tool library

29 open, licence-checked end effectors with files, specs and a simulation status, alongside the families the Studio designs. These ones made it into a recommendation.

Browse the library

Open source, end to end

The same code runs this site, the planner and the physics. Read how it works, or run it yourself.