HansZ8/RoboJuDo

What it solves

RoboJuDo addresses the common problem in robotics research where authors release trained AI models (policies) but provide no accompanying code to actually deploy them on physical hardware or in simulation. It provides a unified, plug-and-play framework that simplifies the process of moving a control policy from a model file to a functioning robot.

How it works

The framework uses a decoupled, modular architecture that separates the control pipeline into three primary components:

• Controller: Handles external input signals, such as those from a keyboard, joystick, or pre-defined motion sequences.
• Environment: Manages the execution interface, whether it is a physics simulator (like Mujoco) or a real humanoid robot (via SDKs).
• Policy: The trained AI model that takes data from the environment and controller to generate specific robot actions.

By keeping these modules separate, users can swap policies or robots without rewriting the core logic. It also supports "multi-policy switching," allowing a robot to transition between different AI behaviors (e.g., switching from a walking policy to a motion-mimicking policy) during a single task.

Who it’s for

RoboJuDo is designed for robotics researchers and developers working with humanoid robots, specifically those using platforms like the Unitree G1, Unitree H1, and FFTAI Gr1T1.

Highlights

• Broad Policy Support: Integrated support for numerous policies including BeyondMimic, ASAP, ProtoMotions, and Unitree official policies.
• Sim-to-Real Pipeline: Built-in workflows to test policies in simulation (Sim2Sim) before deploying them to real hardware (Sim2Real).
• Hardware Agnostic: Supports multiple operating systems (Ubuntu, Windows, macOS) and various humanoid hardware.
• Lightweight Design: Optimized for onboard execution, reducing the need for external cabling (e.g., running on Unitree G1 without Ethernet).