MiniCPM-RobotTrack

MiniCPM-Robot Embodied Series

Real-robot safety: On the first run you must keep runtime.mode: dry-run and must not send any motion commands. With the robot on a stand, an on-site operator should verify the camera, model, latency, speed limits, stop path and the remote-controller/App emergency stop before enabling real control according to the official guide.

Verified Hardware and System

Item	Officially verified configuration
Robot	Unitree Go2 EDU
Compute platform	Jetson Orin NX 16GB
Camera	Go2 front camera; the D435i RGB path must be verified separately
System	Jetson Linux R36.5, JetPack 6
Software stack	CUDA 12.6, TensorRT 10.7, Python 3.10, ROS 2 Humble
Power mode	MAXN mode 0

Data Pipeline

Plain Text

Go2 / D435i RGB
  → TCP JPEG
  → DINO + SigLIP TensorRT
  → MiniCPM-RobotTrack
  → waypoint
  → 限速控制

Deployment Steps

Bash

cd MiniCPM-RobotTrack
python3 scripts/download_upstream_assets.py
python3 -m pip install --user -r requirements-build.txt
./scripts/export_onnx.sh

sudo nvpmodel -m 0
sudo jetson_clocks
./scripts/build_engines.sh
./scripts/preflight.sh
./go2_runtime.py run

Before running, place the complete Hugging Face model snapshot in MiniCPM-RobotTrack/minicpm_robot_track/checkpoints/MiniCPM-RobotTrack/. It should include the custom model code, the tokenizer files and model.safetensors.

Status and Stopping

Bash

./go2_runtime.py status
./go2_runtime.py stop-control
./go2_runtime.py stop

Pre-launch Checklist

• runtime.mode remains dry-run

• Camera input and timestamps are stable

• Model and TensorRT engines load successfully

• End-to-end latency and control frequency meet expectations

• Speed limiting and abnormal-output protection have been verified

• Remote-controller or App emergency stop works

• An on-site operator is in position

Related resources

Topic	Link
MiniCPM-RobotTrack	MiniCPM-RobotTrack
Go2 deployment and reproduction entry point	Go2 deployment and reproduction entry point