FluxVLA Engine: A One-Stop VLA Engineering Platform for Embodied Intelligence

Overview

FluxVLA v0.1.2 expands FluxVLA with end-to-end SARM support, new model families (XVLA and Qwen3-VL), simulation and data-collection tooling (FluxBiSim and FluxDAgger), and broader hardware compatibility including Blackwell GPUs.

Highlights

• Added end-to-end SARM support, including training, manual / VLM-based subtask annotation, and progress inference on LeRobot v2.1 / v3.x datasets, with a dedicated annotation toolkit and RABC weighting utilities.
• Added the XVLA model with a Florence2 backbone and flow-matching head.
• Added Qwen3-VL backbone support (including the Qwen3VL 0.6B + GR00T configuration).
• Added FluxBiSim training and inference support, plus documentation for the FluxBiSim simulation benchmark and the FluxDAgger dual-arm DAgger pipeline.
• Added Blackwell GPU (RTX 5090) compatibility support.
• Added PI and DreamZero configs, and tuned SmolVLA LIBERO finetune hyperparameters.
• Fixed a redundant image resize in the LIBERO eval pipeline and avoided duplicate training progress logs.
• Upgraded to transformers==5.3.0 (existing v0.1.0 environments can upgrade in place without recreating the conda env).

Documentation

This release adds documentation for SARM workflows (docs/sarm.md and tools/sarm_annotate/README.md), refreshes the README with a Performance benchmark table and updated Latest News, adds a Contributing guide, and documents the transformers upgrade path for existing installations.