jev-robotics-demo
MuJoCo Franka arm stacking cubes: code proposes and simulates small moves, Jev picks moves, grasp, release and completion, vs Claude.
Overview
Jev vs LLM: robot arm control drives a simulated Franka arm with a four-finger Allegro hand in MuJoCo to stack a blue cube on a red one, using two approaches. The LLM agent (Claude Opus 5 by default) gets three low-level tools, move_hand_to, set_hand and get_state, and reasons through every step. The Jev agent follows the pattern from jev_fsd: code proposes small moves and simulates each one ahead in a scratch copy of the physics, and Jev picks the move, decides when to grasp and release, and decides when the task is done.
One recorded run each: both stacked the cube, with the Jev agent taking 19.1 s and $0.0006 versus 158.8 s and $0.75 for Opus (which needed one retry; clean Opus runs took about 55 s and $0.19). Times include real-speed arm motion. Any OpenRouter model with tool support can be swapped in for the LLM agent, and runs can be recorded headless and rendered to video.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which proposed move should the hand make next? | — |
| Yes / no | Should the hand grasp / release now? | yes · no |
| Yes / no | Is the stacking task done? | yes · no |
How it works
- Code proposes small candidate moves
- Each move is simulated ahead in a scratch physics copy
- Jev picks the move and grasp or release timing
- Jev decides when the stack is complete
Strengths
- Code handles geometry; Jev handles judgment
- Large time and cost savings in the reported runs
- Easy LLM comparison via OpenRouter
Caveats
- One recorded run per approach
- Simulation only
- Requires TypeSafe and LLM keys
Getting started
Set up Python 3.12 with uv, fetch the needed MuJoCo Menagerie models, export TYPESAFE_API_KEY, then .venv/bin/mjpython jev_agent.py "$TASK"At a glance
- Type
- App
- Status
- prototype
- Language
- Python
- Runtime
- MuJoCo simulation
- Has code
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
A MuJoCo arm demo where local code proposes candidate moves and Jev chooses the target, grasp or release, and completion.
logicrw/awesome-jev-projectsJev (TypeSafe System One) vs Claude Opus 5 driving a simulated robot arm in…
wh000wh000/awesome-jev-live
Frequently asked questions
What does Jev decide?
Which simulated move to take, when to grasp and release, and when the task is done.
How did it compare with Claude Opus 5?
In one run each, Jev took 19.1 s and $0.0006 versus 158.8 s and $0.75.
Can I use another LLM?
Yes, any OpenRouter model with tool support.
Related projects
MuJoCo pilot where Jev, Claude Haiku and reactive rules choose pick-and-place actions; Jev matched the rules exactly.
ProjectresearchPythonFixed-wing landing simulator and benchmark where Jev maps four Choice decisions to aircraft controls against humans, baselines and LLMs.
DemoprototypePythonSimulated 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.
ProjectresearchPython