Skip to content

README.md

Latest commit

 

History

History
93 lines (62 loc) · 3.38 KB

README.md

File metadata and controls

93 lines (62 loc) · 3.38 KB

AgileX Data Processing, Fine-tuning and Inference Guide

This README provides instructions for processing data collected from AgileX robots, fine-tuning GO-1 model, and performing real robot inference.

Table of Contents

Data Processing

1. Analyze Outliers

python agilex_dataset_filter.py --root_dir /path/to/your/hdf5/files

This tool performs comprehensive outlier analysis on HDF5 files by calculating Z-scores based on global mean and standard deviation across all data. It identifies data points that exceed the specified threshold (default 5 standard deviations) and outputs a detailed list of files containing these outliers. The analysis helps ensure data quality by flagging potentially problematic episodes that may contain sensor errors or abnormal robot behaviors.

Arguments:

  • root_dir - Directory containing HDF5 files
  • std_threshold - Standard deviation threshold (default: 5.0)

2. Configure Dataset Features

Set data key names and shapes in agilex_features.py.

3. Convert HDF5 to LeRobot Dataset

# Activate the GO-1 environment
conda activate go1

python convert_agilex_data_to_lerobot.py \
    --src_path /path/to/your/hdf5/files \
    --tgt_path /path/to/output \
    --repo_ids <REPO_IDS> \
    --save_repoid <SAVE_REPO_ID>

Arguments:

  • src_path - Path to the directory containing HDF5 files
  • tgt_path - Path to save the converted LeRobot dataset
  • repo_ids - List of folders to process in src_path
  • save_repoid - Repo ID for saving the converted dataset (default: the first item in repo_ids)

Notes:

  • We use the observations/qpos as both state and action. (The default action key cannot replay.)

  • Set the exclude_files list in the script to exclude files containing outliers.

Model Fine-tuning

Refer to the GO-1 Fine-tuning Guide for detailed instructions.

Model Deployment

1. Start GO-1 Server

Start the GO-1 inference server using your fine-tuned model checkpoint and data statistics:

conda activate go1

python evaluate/deploy.py --model_path /path/to/your/checkpoint --data_stats_path /path/to/your/dataset_stats.json --port <SERVER_PORT>

The server will will listen on port SERVER_PORT and wait for observations.

2. Start AgileX Client

python agilex_inference.py --host <SERVER_IP> --port <SERVER_PORT> --ctrl_type joint --publish_rate 30 --chunk_size 30

Arguments:

  • host - IP address of the server (default: 127.0.0.1)
  • port - Port number of the server (default: 9000)
  • ctrl_type - Control type of the robot (joint or eef, default: joint)
  • publish_rate - Action publishing frequency (default: 30)
  • chunk_size - Executed action chunk size, should be smaller than the action chunk size of the model (default: 30)