MiniCPM-RobotManip

MiniCPM-Robot Embodied Series

This page focuses on local inference and deploying the evaluation service. Before integrating into production, first use offline samples to verify input preprocessing, action dimensions and control limits.

Deployment options

Option	Use case
Single-process inference	Local sample verification, algorithm debugging and integration testing
Model service	Decouples the model from standalone simulation environments such as LIBERO, CALVIN, RoboTwin2 and RMBench

Base environment

• Python 3.10
• PyTorch 2.6.0
• CUDA 12.4
• Full MiniCPM-RobotManip model weights

Bash

cd MiniCPM-Robot/MiniCPM-RobotManip
conda env create -f environment.yml
conda activate MiniCPM-RobotManip

Local inference checks

1. Confirm that images can be read and that their color space and resolution meet the model's preprocessing requirements.
2. Confirm that the language instruction matches the current task.
3. Provide the correct state file and embodiment ID for your robot.
4. Check the shape and value range of the output action chunk before connecting the downstream controller.

Model service and simulators

The official evaluation connects to the model service over WebSocket and MessagePack. The model service and each benchmark simulator run in separate Python environments; the simulators, datasets and assets must be installed separately according to the corresponding evaluation guide.

Deployment checklist

• Model weights and custom code load successfully

• Both single-image and multi-view inputs have been verified

• Robot state dimensions and embodiment ID have been confirmed

• Action outputs pass clipping, timeout and outlier checks

• Verified in simulation or offline replay before connecting to the real robot

Related resources

Topic	Link
MiniCPM-RobotManip	MiniCPM-RobotManip
Evaluation directory	Evaluation directory