Skip to content
Effector Lab

Tutorial · 10 minutes

Plan a task

The Studio starts from a job in plain words. The planner turns it into objects and steps the simulator can run; you check and edit them, then download the plan to run the search on it.

01What a plan is

A plan has two lists. Objects are parametric templates the simulator knows: a part, a bin, a hinged lid or door, a drawer, a latch, a threaded joint, a valve, a port. Steps are skills on those objects: grasp, suction, wrap, hook pull, twist, insert, push. Every step carries the test that says it is done and the limits it must respect.

The plan behind the returns tote example, as the simulator reads it:

1 Objects (4)

  • ToteBin or tote · height 120 mm, length 300 mm, wall 4 mm
  • Tote lidHinged lid or door · damping 0.05 N m s, depth 230 mm, feature 15 mm
  • Small cartonRigid part · mass 300 g, size x 120 mm, size y 80 mm
  • Shrink-wrapped packRigid part · mass 150 g, size x 150 mm, size y 100 mm

2 Steps (4)

  1. 1Open the lidHook pullRigidDone when angle 95 to 110 deg · at most 20 N
  2. 2Pick the cartonGraspRigidDone when ends inside the target; tilt at most 30 deg
  3. 3Pick the packGraspRigidDone when ends inside the target; tilt at most 30 deg
  4. 4Close the lidPushRigidDone when angle 0 to 5 deg · at most 20 N

3 Every step names a skill, the object it acts on, when it counts as done, and its limits.

02Describe the job

Open the Studio and write what the robot has to do, or start from an example. The planner reads what a technician would say: each object with its size and mass, what happens in what order, and the limits.

The example's description; the highlighted numbers become parameters and limits
Warehouse returns: a plastic tote (300 x 200 mm, 120 mm tall) stands on the packing bench with its hinged lid shut; the lid overhangs the tote by a 15 mm lip all round. Open the lid, put a small rigid carton (about 120 x 80 x 60 mm, 300 g) and a flat shrink-wrapped pack (about 150 x 100 x 20 mm, 150 g) from the bench into the tote, then close the lid.
  • Name the objects

    Say what each thing is and how big and heavy it is: a 120 x 80 x 60 mm carton of 300 g, not a box.

  • Say the order

    One sentence per action, in order. The planner keeps your order as step dependencies.

  • Give the limits

    A force the lid can take, a torque to tighten to, a precision. Limits become checks on every run.

03Press Plan

The Task Engineer (NVIDIA Nemotron 3 Super) answers with one forced tool call: the objects as templates and the steps as skills, with bounded parameters. Code then checks every rule. A plan that breaks one goes back once with the list of problems.

The example: valid after the layout fixModel calls: 2Layout fixes: 2

Explore mode on this site

Plan replays the recorded run of each example, model calls included, so you see what the planner did. To plan your own words live, use your own key or a small free tier in the “Plan and run it live” panel, or run the Studio locally with your own Token Factory key (Run it yourself), or press “Build a plan by hand” and fill in the objects and steps yourself.

04Check and edit

Every object and step can be edited, added, moved or removed, and the plan is checked again on each change: the line above the lists says whether it is valid. Look at three things before you run it.

  • Fidelity

    Each step says how it is simulated:

    • RigidA rigid multibody model with a scripted controller.
    • ApproximationA labelled approximation (suction seal, soft arm, latches, threads, valves): verify on hardware.
    • Not simulated yetNo physics model yet: never scored as a pass.
  • Assumptions and questions

    What the planner assumed, and what it would ask you. An answer is assumed until you change the plan. The example made 6, for instance: “The lid overhangs the tote by a 15 mm lip on all sides.”

  • What cannot be simulated

    Cloth, liquids, food and other soft things are out of scope. The planner lists them with the nearest skill it can run, and a not simulated yet step never counts as a pass.

05Download the plan

When the plan is valid, Download TaskSpec JSON saves it as taskspec-<id>.json (or Copy JSON puts it on the clipboard). That file is the input of the search: every tool tried on every step, the combinations, the finalists and the build pack.

06Plan and run it on this site

The Studio can plan your own words live and then test the plan in physics, with nothing to install. In the panel called “Plan and run it live”, choose who pays for the planning call:

  • Your NVIDIA API keyNemotron plans on your own account.
  • Your Nebius Token Factory keyNemotron plans on your own account there.
  • The free allowanceA daily allowance of plans that every visitor shares. The panel counts down what is left.

Your key

Your key is sent once over HTTPS for this request and is never stored or logged. It stays in the page's memory only, and it is forgotten after each request unless you untick the box.

The hosted runner is a free trial with daily limits: each visitor gets 3 free plans and 2 physics runs a day (up to 30 plans with your own key), from a small shared budget. Everything it does also runs on your own computer, without limits.

Press Plan live. Planning usually takes 15 to 45 seconds and can take up to five minutes. The plan loads into the steps below the panel exactly as a recorded one does, so you can check and edit it. When the plan is valid, press Run in physics. The run goes to a hosted runner, which can take up to three minutes to accept a plan it has not seen before. The address gets a #run- token, so reloading the page picks the run back up.

Runs on Nebius Serverless Jobs, capped at $2 a day, and the free allowance has its own daily cap. When a cap is reached, the page says which one and when it resets. If the runner is busy, the page says to try again in a minute. If the hosted runner is off, the page says so and points to Run it yourself.

A finished run shows a summary card: the recommended setup and how often it completed the task, the best single tool beside it, a table of the steps, and a link to the full report. Replays are shown for the recorded examples; hosted runs show the summary. Simulated, verify on hardware. Product mode: not part of the pre-registered experiment.