EmbodiedJev: Open-Source Jev-Powered Robot Arm Simulation for Pick, Stack, Avoid

Project Overview

EmbodiedJev (Chinese name: 行知) is an open-source embodied intelligence experiment that connects the Jev fast decision-making model — and other large language models — to a robot simulation environment. The system lets AI control a simulated Franka Panda robotic arm in the browser to perform manipulation tasks without requiring physical hardware or pre-trained robot policies.

The simulation runs on the MuJoCo physics engine, so every grasp, lift, and placement is validated by real-time physics rather than being a pre-recorded animation.

Three Benchmark Tasks

Pick-and-place : grasp a red cube and place it into a blue tray.

Block stacking : place one cube on top of another and verify stability.

Obstacle-avoidance placement : lift the object, clear an obstacle, then complete the placement.

Together these tasks cover the core manipulation primitives: grasping, moving, obstacle avoidance, and placing.

How Jev Controls the Arm

Jev is a decision model, not a dedicated robot policy. EmbodiedJev bridges this gap with a perception–decision–action loop :

The simulator extracts the current state (arm pose, cube positions, gripper status) and serializes it as structured text — Jev does not yet consume raw camera images.

The model selects the next action (e.g., move toward cube, adjust pose, close gripper).

A low-level controller executes the action in MuJoCo and returns the updated state.

The loop repeats until the task succeeds or a step limit is reached.

This step-by-step, closed-loop approach lets the model react to physical feedback (slippage, collisions) just like a human operator would.

Three Control Modes

Preset skills : the model chooses from a library of pre-defined motion primitives.

Stepwise Cartesian control : the model directly commands ΔX, ΔY, ΔZ movements and gripper open/close.

Hierarchical decision-making : the model first sets a sub-goal (approach, grasp, transport) then picks the concrete action. This separation of high-level planning from low-level control makes model behavior easier to inspect.

Multi-Model Support

The framework is model-agnostic. Developers can plug in Jev, GPT, Claude, or a local MiniCPM5-2B and run identical tasks to compare decision speed, action quality, and API cost. A rule-based baseline is also provided for zero-API-key experimentation.

Quick Start

git clone https://github.com/FBddcz/embodied-jev.git
cd embodied-jev

macOS / Linux:

python3 -m venv .venv
source .venv/bin/activate
python -m pip install -e .
npm ci
npm run build
embodied-jev serve --port 8090

Windows PowerShell:

py -3.11 -m venv .venv
.venv\Scripts\Activate.ps1
python -m pip install -e .
npm ci
npm run build
embodied-jev serve --port 8090

Then open http://127.0.0.1:8090 in a browser to run the first experiment.

Limitations

EmbodiedJev remains a simulation-only research prototype . It demonstrates that general-purpose decision models can drive a physics-backed manipulator in closed loop, but does not yet transfer to real-world robots (sim-to-real gap, latency, sensor noise, etc. are not addressed).