jev-libero
Fine-grained robot control on LIBERO tasks where Jev picks intent, motion family and input with local physics previews.
Overview
jev-libero explores robot manipulation with Jev on LIBERO tasks. Jev makes layered decisions, selecting an intent, then a contact or motion family, then one of 27 low-level inputs (Cartesian translations, wrist rotations, gripper open or close, hold), each choice informing the next. Before execution, reversible simulator branches preview candidate effects locally. Tasks are configured in JSON, selecting measurements, exported features, contact rules, goals and per-layer Jev inputs through one interface.
Recorded demos include closing a microwave (14 decisions, 111 environment steps), closing a top drawer (20 decisions, 155 steps) and grasping and lowering a soup can into a basket (40 decisions, 314 steps). Runs save requests, predictions, controls, simulator states, costs and video together, and an interactive Decision Lab replays Jev's choices and probabilities beside each video. It needs Python 3.10 or 3.11 and a LIBERO, robosuite and MuJoCo environment to run episodes. Documentation is in English and Chinese.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | What is the current intent? | — |
| Choice | Which contact or motion family applies? | — |
| Choice | Which of 27 control inputs should be applied? | — |
How it works
- Load a JSON task definition
- Jev chooses intent, motion family and input in layers
- Preview candidate effects in reversible simulator branches
- Execute and record everything
Strengths
- Fine-grained control from a decision model
- Physics previews before acting
- Inspectable recorded runs
Caveats
- Simulation only
- Requires a LIBERO environment to run
Getting started
`git clone https://github.com/Dimweaker/jev-libero.git && cd jev-libero && pip install -e . && jev-libero tasks`.At a glance
- Type
- App
- Status
- research
- Language
- Python
- Runtime
- LIBERO / robosuite / MuJoCo
- Has code
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
Two LIBERO tasks, one control engine. Each demo loads its own JSON task definition. Videos follow simulation time, with decision and physics-preview waiting omitted.
logicrw/awesome-jev-projectsFine-grained robot control with Jev, physics previews, and configurable LIBERO tasks.
hellogumbo/awesome-jevFine-grained robot control on LIBERO tasks with physics previews and configurable task definitions.
MrJev/awesome-jev
Frequently asked questions
What does Jev decide?
Intent, motion family and low-level control input at each step.
How are mistakes avoided?
Candidate effects are previewed in reversible simulator branches.
Can I watch runs?
Yes, in the interactive Decision Lab.
Related projects
Simulated Franka arm that follows plain-English goals by having Jev chain about 30 hardcoded motion primitives, one per step.
Agent skillprototypePythonCamera-only MuJoCo quadrotor where Jev makes 2.5 Hz tactical judgments while code keeps flight control and a safety veto.
ProjectprototypePythonMulti-drone simulation where Jev makes fast reflex decisions per drone and an optional slower LLM may advise but never takes control.
AppprototypeTypeScriptOpen 4B System I driving model that scores behaviours like cruise, stop, yield and brake from three camera frames each decision.
ProjectresearchPythonExperimental robot loop where Jev chooses the next verified action from perception-generated options for a language goal.
ProjectresearchPythonMuJoCo PiPER arm controlled by a local decider-2b model choosing among ten skills via a Jev-style typed-choice interface while targets move.
Agent skillresearchPython