FluxVLA/FluxVLA

FluxVLA Engine – A One‑Stop Platform for Embodied‑Intelligence Research

What it is – FluxVLA Engine is an open‑source, end‑to‑end software stack that lets researchers and engineers train, evaluate, and deploy vision‑language‑action (VLA) models for robots. It bundles:

• Unified configuration and standardized APIs so the same code can run in simulation, on a workstation, or on edge devices.
• Modular components (data pipelines, model back‑ends, reward‑model trainers, inference runtimes) that can be swapped without rewriting the whole system.
• Deployability – Docker images, Jetson‑Orin support, Triton‑backed inference, and ZMQ remote‑inference wrappers.

Why it matters – Building robot‑ready VLA models usually requires stitching together many disparate tools (simulators, large‑scale model training, hardware drivers). FluxVLA removes that friction, providing a single “engineering loop” from raw data to real‑robot execution, which lowers the barrier for both academia and industry to experiment with embodied AI.

---

Key Features (as described in the README)

Feature	What the README says
Full‑stack platform	Handles everything from data collection, model training, benchmark evaluation (LIBERO, RoboCasa, RoboDojo) to real‑robot deployment (Franka, Oli humanoid, Jetson Orin).
Unified configuration	One‑command installer (scripts/install_env.sh) with modes for simulation‑only, real‑robot‑only, or full stack.
Standardized interfaces	Same Python APIs work across simulators and hardware; remote inference via ZMQ; Triton backend for accelerated inference.
Module decoupling	Separate model families (SmolVLA, Cosmos3‑Edge, GR00T N1.5/N1.7, Qwen3VL, DreamZero, PI0/PI0.5, FastWAM, DiT4DiT) can be swapped.
Deployability	Docker images, Jetson‑Orin flashing guide, RTX‑5090 real‑time inference (45 Hz), edge inference (7.4 Hz).
Benchmark suite	Pre‑trained checkpoints and scripts for LIBERO, RoboCasa GR1, RoboDojo, with detailed performance tables.
Extensible pipelines	FluxDAgger (model‑decoupled DAgger), FluxBisim (simulation benchmark), reward‑model reweighting (RA‑BC/AW‑BC).
Documentation	Multi‑language docs (EN/中文/日本語), API reference, deployment guides, and a technical report on arXiv.

---

Performance Highlights (from the README tables)

• LIBERO benchmark – Best average scores around 98 % (e.g., FluxVLA (PI0.5) 97.95 % average, FluxVLA (DiT4DiT) 98.65 %).
• RoboCasa GR1 – Top model FluxVLA (DiT4DiT) reaches 57.25 % average success across cabinet, drawer, microwave, and generalization tasks.
• RoboDojo – FluxVLA (PI0.5) shows the highest average progress score 13.61 and success 8.83 %, indicating strong long‑horizon and memory capabilities.

---

Getting Started (Installation)

# Create a fresh conda env
conda create -n fluxvla python=3.10 -y
conda activate fluxvla

# One‑command installer – pick a mode
bash scripts/install_env.sh sim-only   # simulation & benchmark only
# bash scripts/install_env.sh real-only  # real‑robot drivers & remote inference
# bash scripts/install_env.sh full       # everything

The installer automatically selects the appropriate CUDA‑PyTorch build and pulls RoboCasa assets from HuggingFace unless you skip them.

---

Typical Workflow

1. Configure a YAML/JSON experiment file (environment, model, reward model).
2. Train using the provided scripts (scripts/train.sh) – supports large‑scale models (Cosmos3‑Edge, GR00T, DiT4DiT) and checkpoint‑free GPT‑6 Astra evaluation.
3. Evaluate on standard suites (LIBERO, RoboCasa, RoboDojo) with one‑line commands that log progress and success rates.
4. Deploy – build a Docker image, flash a Jetson Orin, or run remote inference via ZMQ; control real robots (Franka, Oli) using the same model checkpoint.

---

Who Might Use It

• Robotics researchers needing a reproducible pipeline for VLA experiments.
• AI labs that want to benchmark new vision‑language‑action architectures against state‑of‑the‑art baselines.
• Industry teams building embodied AI products (warehouse pick‑and‑place, home‑assistant robots) who need a ready‑made deployment stack.

---

Where to Find More

• Code & checkpoints – https://huggingface.co/limxdynamics/FluxVLAEngine
• Technical report – https://arxiv.org/abs/2609.17210
• Full documentation – https://fluxvla.limxdynamics.com (English, Chinese, Japanese)
• Community – WeChat and Feishu links in the badge bar for support and announcements.

---

Bottom line – FluxVLA Engine is a genuine, comprehensive open‑source platform for training and deploying vision‑language‑action models in embodied‑intelligence settings, complete with benchmarks, real‑robot demos, and extensive deployment tooling.