This pull request introduces comprehensive documentation and example code for training and deploying StarVLA with Franka robot data, including both single-arm and dual-arm setups. The main additions are a detailed step-by-step guide in README.md and a fully annotated dual-arm inference example script. These changes make it much easier to understand the workflow for data preparation, model training, and real-time robot control, and provide reusable code for WebSocket-based policy server integration.

Documentation improvements:

• Added a detailed guide to examples/Franka/README.md covering data conversion, dataset registration, training configuration, model parameter setup, deployment, and adaptation to new robots. The guide includes code snippets, configuration examples, and reference implementations for env.step() in both single-arm and dual-arm setups.

Example code enhancements:

• Added examples/Franka/eval_files/inference_dual_example.py, a dual-arm inference script with clear separation between reusable real code (WebSocket client, request/response handling, action unnormalization, normalization stats loading) and robot-specific pseudocode (camera acquisition, robot environment, action execution). The script is annotated for easy adaptation to other robots and includes detailed comments on action space and integration.

Deployment and integration guidance:

• Provided explicit instructions and reference implementations for adapting StarVLA to new robots, including how to handle action space differences, normalization statistics, and robot-specific interfaces.

Action normalization and unnormalization logic:

• Included robust, reusable code and documentation for mapping normalized model outputs to real robot actions, supporting both single-arm and dual-arm action spaces, with clear handling of gripper binarization and dimension masks.

WebSocket-based policy server workflow:

• Clarified the architecture and workflow for client-server communication, including request formats, image handling, and action execution, enabling straightforward integration with real robot systems.