Unitree R1 Edu embodied AI

Task: Replace Existing Onboard LLM with Local GPT-2B

Hardware

* Unitree R1 EDU
* Onboard NVIDIA Jetson Orin NX / onboard GPU compute platform (~100 TOPS)
* Existing robot software and motor-control system are already working.
* The robot currently has a built-in Chinese AI chatbot / LLM. The exact model is not yet confirmed and is not Qwen.

Main Objective

Replace the robot’s existing onboard LLM/chatbot with a local GPT-2B-class language model running directly on the robot’s onboard GPU.

The new local model must perform the same high-level conversational / command-understanding function as the existing chatbot, while the existing robot control and motor-control system must remain unchanged.

In other words:

Existing system:

User Voice/Text → Existing Chinese LLM → Existing Robot Control Interface → Motors

Target system:

User Voice/Text → Local GPT-2B → Existing Robot Control Interface → Motors

The GPT-2B model should effectively replace the existing LLM layer, not the robot’s control layer.

What the freelancer needs to do

1. Identify the current AI/LLM architecture
* Determine what model currently powers the onboard Chinese chatbot.
* Identify how the current LLM communicates with the robot’s existing control system.
* Identify the existing API, SDK, ROS interface, middleware, or function-calling mechanism used between the LLM and robot control.
2. Deploy a local GPT-2B model
* Select an appropriate ~2B parameter model suitable for local inference.
* Deploy it on the robot’s onboard NVIDIA GPU.
* Use GPU-accelerated inference (TensorRT-LLM, llama.cpp CUDA, vLLM, or another appropriate optimized runtime).
* Optimize the model for low latency and reasonable memory/compute usage.
3. Connect GPT-2B to the existing robot interface
* Do NOT rewrite the existing motor-control system.
* Do NOT develop new low-level motor drivers.
* Do NOT replace the existing Unitree SDK / motion-control implementation.
* Instead, make the new LLM communicate with the existing high-level robot-control interface in the same way the current chatbot does.
4. Function / command calling

The local GPT-2B should be able to interpret natural-language commands and generate structured commands/functions that are already supported by the robot.

For example:

"Stand up and walk forward"

→ GPT-2B interprets intent

→ calls existing robot functions such as:

stand()

walk_forward()

The exact available functions/interfaces should be determined from the existing R1 EDU software.

5. Voice interface

If the existing system already provides speech recognition, preserve it.

If necessary, the architecture should support:

Microphone → Local STT → GPT-2B → Existing Robot Functions

The primary task is LLM replacement, not development of a new speech-recognition system.

Important constraint

The freelancer must reuse the existing robot-control stack.

The project is NOT:

* rewriting robot locomotion
* rewriting motor control
* developing a new Unitree SDK layer
* developing a new walking controller
* developing a new ROS control architecture

The project IS:

Replacing the existing onboard Chinese LLM with a locally running GPT-2B model and connecting the new model to the existing robot-control interface.

Expected final architecture

Microphone / Text Input
↓
Existing or Local STT
↓
Local GPT-2B running on onboard NVIDIA GPU
↓
Structured command / function call
↓
Existing R1 EDU Robot API / ROS / SDK
↓
Existing Motion & Motor Control
↓
Robot

Deliverables

1. Local GPT-2B model running on R1 EDU onboard computer.
2. GPU-accelerated inference.
3. Integration with the existing robot-control interface.
4. Natural-language command → structured robot function execution.
5. Source code for the integration layer.
6. Installation/deployment instructions.
7. Ability to run the system autonomously on the robot without cloud LLM/API dependency.
8. Basic demonstration showing that the new GPT-2B successfully replaces the existing chatbot while the existing robot-control system continues to operate unchanged.

Freelancer requirements

Experience with:

* NVIDIA Jetson Orin / embedded GPU
* Local LLM deployment
* 1–3B parameter models
* TensorRT-LLM / llama.cpp / CUDA optimization
* ROS / ROS2
* Unitree robots or similar robotic platforms
* LLM function calling / tool calling
* Python and/or C++

Experience with embodied AI / LLM-to-robot integration is strongly preferred.

Key point

The freelancer should treat the existing chatbot as a replaceable AI module.

We already have the robot’s motor-control and motion-control functionality. The goal is to replace only the LLM/AI reasoning layer with a local GPT-2B model running on the robot’s onboard GPU.