FBddcz/embodied-jev

What it solves

EmbodiedJev (行知) provides a browser-based workbench for experimenting with Embodied AI. It allows users to visualize how AI models make decisions for robotic tasks—such as grasping, stacking, and obstacle avoidance—without requiring physical robots or complex setup. It bridges the gap between high-level model reasoning and low-level physical execution.

How it works

The project uses the MuJoCo physics engine to simulate a Franka Panda robotic arm. Users can connect various AI models (including local models like MiniCPM5-2B or cloud APIs like Claude and OpenAI-compatible services) to the simulation. The system follows a loop: the model observes the state (via structured data or RGB camera input), selects an action (such as a movement step or a skill), and the physics engine executes it, providing feedback for the next decision cycle. It supports side-by-side comparisons of different models under identical conditions.

Who it’s for

This project is for AI researchers, robotics enthusiasts, and developers interested in testing how LLMs and vision-language models perform in physical decision-making tasks. It is designed to be accessible for those with zero robotics experience while offering enough depth for advanced users to perform custom experiments and benchmarks.

Highlights

• Real-time Visualization: Watch the robot's decision-making process, including stage selection, XYZ movement, and gripper actions.
• Model Agnostic: Supports local models (MiniCPM) and various cloud-based APIs (OpenAI, Anthropic, TypeSafe Jev).
• Comparative Analysis: Run multiple models simultaneously on the same task and seed to compare performance, latency, and costs.
• Physics-based Feedback: Uses MuJoCo for realistic interaction, including grasping, collision detection, and object manipulation.
• Extensible: Allows users to define custom tasks, import/export experiment data, and adjust camera configurations.