X-VLA M2W Multitask Fine-tuning

This repository contains the final checkpoint, training configuration, logs, and reproduction scripts for fine-tuning X-VLA on a real-world M2W multitask robot dataset.

中文说明：本仓库提供 X-VLA 在 M2W 多任务真实机器人数据上的微调产物，包括最终 checkpoint、训练配置、训练日志和数据处理/训练/验证脚本。

Final checkpoint

The delivered checkpoint is:

checkpoints/020000/

• Training steps: 20,000
• Validation result: eval_loss = 0.0180 at the final step
• Checkpoint integrity: validated after training; model, processors, optimizer, scheduler, and RNG state are present.

Data

The training data combines two M2W tasks:

• table_clean from yuuu94/M2W-VLA-table-clean-robotwin
• put_mango from yuuu94/Real-World-M2W-Demo-Episodes-put-mango

The merged training dataset contains 200 episodes and 120,469 frames, with three camera views and 14-dimensional robot state/action data. The raw dataset is not redistributed in this repository.

Base model

The starting policy was:

• yuuu94/realworld_put_mango_stride2_h16_implicit_cot_phs_bs48_warmup1k_pyav

Training setup

The primary configuration is available at:

training/xvla_full_1gpu.yaml

Key settings:

• X-VLA / LeRobot training stack
• One GPU, bfloat16
• Batch size: 16
• Optimizer learning rate: 1e-4
• Warmup steps: 1,000
• Total steps: 20,000
• Checkpoint interval: 5,000 steps
• Random seed: 42

Repository layout

checkpoints/020000/       Final complete checkpoint
training/                 Training configuration and log
*.py                      Data preparation, training, resume, and validation scripts

Environment

Install a CUDA-enabled PyTorch build first, then install the dependencies:

pip install -r requirements-train.txt

The experiment used lerobot[dataset,training,xvla]==0.6.0 and PyAV for video decoding.

Notes

• This repository is intended for research and reproduction.
• The checkpoint is a training checkpoint and includes optimizer/scheduler/RNG state, so it can be resumed.
• Please evaluate the policy on the target robot setup before any real-world deployment.