Community · E2 · artifact verified

Control a robot arm through layered choices

Jev drives a LIBERO robot through 27 control inputs with layered decisions - intent, then motion family, then input - while reversible physics previews evaluate candidate effects locally before anything executes.

01 · Role in the system

What Jev does here

Three layers narrow the decision: Jev selects an intent, then a contact and motion family informed by that choice, then one of 27 inputs covering Cartesian translations, wrist rotations, gripper open and close, and hold. Before an input executes, reversible simulator branches preview candidate effects locally, so each choice is made against predicted physics rather than hope. Everything is inspectable: model requests, predictions, controls, simulator states, costs, and trajectory media save together, and an interactive Decision Lab replays every run with Jev's choices and probabilities alongside the video. Three task configurations share one engine - closing a microwave in 14 decisions over 111 environment steps, a drawer in 20 over 155, and grasping and lowering a soup can in 40 over 314 - with task measurement, features, contact rules, and per-layer inputs configurable in JSON.

Twenty-seven inputs, three layers, physics before commitment

The decision architecture narrows in three layers — intent, then contact and motion family, then one of 27 inputs spanning Cartesian translations, wrist rotations, and gripper commands — and before any input executes, reversible simulator branches preview its candidate effects locally. Choices are made against predicted physics rather than hope, which is the property physical robotics always needs and simulations rarely show this explicitly.

Everything persists together: model requests, predictions, controls, simulator states, costs, and trajectory media, replayable in an interactive Decision Lab that shows the choices and probabilities beside the video. Three task configurations run on one engine — a microwave closed in 14 decisions over 111 steps, a drawer in 20 over 155, a soup can grasped and lowered in 40 over 314 — all configurable in JSON.

Simulation results, labeled as such

The record is explicit that this is the LIBERO benchmark environment, not physical hardware, with three curated task configurations and chosen seeds rather than a leaderboard claim. Success counts are the author's recorded runs — and because the records ship with the repository, each one is replayable rather than taken on faith.

As robotics-cluster evidence, this is the most inspectable control record in the index: the layered choice space bounds what the model can express, the physics preview bounds what it can cause, and the replay machinery makes both bounds visible to a reviewer in minutes.

02 · Control boundary

Where Jev sits

Layered choices over a fixed 27-input space with local reversible physics previews before execution; every run persists request, control, state, and media for replay.

Code owns the loop, permissions, thresholds, validation, and side effects. Jev owns only the bounded judgments described above.

03 · Known limits

What this evidence does not prove

  • Simulation only: LIBERO benchmark tasks, not physical hardware; the videos follow simulation time with waiting omitted.
  • Three curated task configurations with chosen seeds, not a benchmark leaderboard claim.
  • Success counts are the author's recorded runs, replayable from the shipped records.

04 · Attribution

Public sources

This is a Community record: the project was published by a third-party community author.