# FluxVLA FAQ

## General Features

**Q: What is FluxVLA?**

A: FluxVLA is an open-source, one-stop VLA (Vision-Language-Action) engineering platform from LimX Dynamics, covering the full engineering loop from data processing and model training to simulation evaluation and real-robot deployment. Built on four core design principles — unified configuration, standardized interfaces, module decoupling, and deployability — it manages entire experiments through a single config file, significantly lowering the engineering barrier for VLA research and reproduction. The technical report is available on arXiv (2609.17210), with code, model checkpoints, and training datasets fully open-sourced.

---

**Q: Which models does FluxVLA support?**

A: FluxVLA provides FluxVLA-native integrations of mainstream open-source VLAs and world-action models, including Pi0, Pi0.5, GR00T N1.5 / N1.7, SmolVLA, OpenVLA, LlavaVLA, DreamZero, DiT4DiT (built on Cosmos-Predict2.5), Cosmos3 (Nano / Super / Edge), and FastWAM (uncond / Joint / IDM variants). All models come with unified training and evaluation configs, and support modular swapping of LLM backbones (Llama / Gemma / Qwen families), vision backbones (DINOv2 / SigLIP), and VLM backbones (PaliGemma / Qwen-VL).

---

**Q: Beyond model integrations, what key capabilities does FluxVLA offer?**

A: Five representative capabilities: 

① closed-source model evaluation — checkpoint-free GPT-6 Astra evaluation on LIBERO and RoboCasa via the OpenAI Responses API with multi-view observations; 

② large-scale parallel evaluation — a FluxThemis ZMQ multi-GPU evaluation server supporting benchmarks such as RoboDojo; 

③ fast adoption of new models — native support for GR00T N1.7, DiT4DiT, Cosmos3, and FastWAM across simulation and Franka real robots; 

④ complete training recipes — RoboCasa full-data recipes for eight models with open-sourced checkpoints; 

⑤ humanoid whole-body control — loco-manipulation inference for the Oli humanoid with per-finger MROS hand control and configurable inference modes. 

Detailed per-version changes are documented in the GitHub CHANGELOG.

---

## Training & Evaluation

**Q: Which simulation benchmarks does FluxVLA support, and how does it perform?**

A: FluxVLA currently supports three simulation benchmarks — LIBERO, RoboCasa GR1, and RoboDojo — with officially reproduced results and open-sourced checkpoints. On LIBERO, FluxVLA(DiT4DiT) achieves a 98.65 average success rate and FluxVLA(Pi0.5) reaches 97.95. On RoboCasa GR1 with full-data training, FluxVLA(DiT4DiT) leads at 57.25% average success (50 trials per task). RoboDojo covers 42 benchmark tasks with 2,100 trials per model, reported as progress score / success rate. All results are verifiable in the README performance tables, with checkpoints released on Hugging Face.

---

**Q: How can I evaluate closed-source models like GPT-6 with FluxVLA?**

A: FluxVLA supports checkpoint-free evaluation of GPT-6 Astra through the OpenAI Responses API: the evaluation script directly calls the API policy with multi-view observations and native LIBERO action control. API credentials are configured via environment variables, and rollout artifacts are automatically organized. This capability covers the full LIBERO suite and RoboCasa GR1, enabling quick assessment of closed-source models' zero-shot performance on embodied tasks.

---

**Q: What data format is required? Can I use my own data?**

A: FluxVLA standardizes on the LeRobot Dataset format (v2.1 recommended; SARM workflows also support v3.x) with Parquet dataset loading. Self-collected data (e.g., HDF5 files from ALOHA robots) can be converted to LeRobot format following the data conversion guide in the docs. Note: when mixing multiple datasets for training, ensure all data conforms to LeRobot v2.1 — otherwise you will hit feature-alignment errors. We recommend pinning a specific LeRobot commit to guarantee format consistency.

---

**Q: Can I run simulation evaluation on devices without ray tracing (e.g., A100)?**

A: Yes. Although LIBERO and RoboCasa rely on ray tracing for rendering by default, FluxVLA supports GPU rendering via EGL on non-ray-tracing devices. The one-command installer (sim-only / full mode) automatically probes the MuJoCo EGL environment and configures the required system dependencies and activation hooks; manual EGL setup instructions are also available in the documentation FAQ.

---

## Deployment & Inference

**Q: Which real robots does FluxVLA support?**

A: Validated real-robot deployment paths include Franka arms (single / dual, with joint and end-effector pose control), UR3, AgileX ALOHA dual-arm, the LimX Dynamics Tron2 humanoid, and a whole-body (loco-manipulation) inference path for the Oli humanoid — supporting sensor input, WebSocket control, per-finger MROS dexterous hand control, and keyboard-pause / continuous inference modes. Deployment guides and example configs are available in the docs directory.

---

**Q: How fast is inference? Are there acceleration options?**

A: FluxVLA offers multi-level inference acceleration: 

① GR00T-RTC accelerated inference reaches 45 Hz on an RTX 5090; 

② Pi0.5-RTC supports a Triton backend with fused Triton kernels, CUDA Graph capture, and custom CUDA operators; 

③ edge deployment on NVIDIA Jetson Orin is supported (official Docker image provided), with GR00T-N1.5 reaching 7.4 Hz on-device; 

④ RTC (Real-Time Chunking) is supported to improve cross-chunk trajectory continuity. 

See the inference acceleration documentation for details.

---

**Q: How can resource-constrained edge devices run large-model inference?**

A: FluxVLA provides a ZMQ-based remote inference framework (server / client architecture) that offloads model inference to remote GPUs, while the edge device only handles sensor data upload and action execution. A FluxThemis ZMQ evaluation server was also added recently, supporting multi-GPU replicas, episode-affinity scheduling, and configurable parallel worker startup — well suited for large-scale parallel evaluation and multi-machine deployment.

---

## Installation & Community

**Q: What are the most common installation pitfalls?**

A: Three frequent issues: 

① transformers version — FluxVLA does not pin transformers in requirements.txt; install it manually (4.53.0 recommended for Pi0 / GR00T / SmolVLA, 4.40.1 for OpenVLA, 4.53.2 for Tron2 deployment; use separate Conda environments when working with multiple models); 

② NumPy version being overwritten — pin it with `pip install numpy==1.26.4` after installation; 

③ egl\_probe build failures — install CMake first (via conda, not pip), and set `CMAKE_POLICY_VERSION_MINIMUM=3.5` if your CMake version is too new. See the documentation FAQ and README FAQ sections for more.

---

**Q: What if I can't connect to Hugging Face when downloading models or datasets?**

A: Set the environment variable to use the hf-mirror endpoint before running the download command: `export HF_ENDPOINT="https://hf-mirror.com"`. The one-command installer and the RoboCasa asset downloader both support this mirror endpoint by default.

---

**Q: How do I debug and develop on top of FluxVLA?**

A: FluxVLA's training and evaluation scripts launch distributed jobs via torchrun, which is incompatible with VSCode's default debugging. The documentation FAQ provides a complete `.vscode/launch.json` template using torchrun as the debug entry program for breakpoint debugging (we recommend setting `nproc_per_node` to 1 and disabling wandb uploads during debugging). For development, please read the contributing guide first: open a GitHub Issue to align on scope for new features / models, branch from upstream/main with prefixes like feat/ or fix/, and make sure pre-commit checks and CI pass before submitting.

---

**Q: Where can I ask questions when I run into problems?**

A: Three channels: 

① the OpenClaw assistant in the upper-right corner of the project documentation site (fluxvla.limxdynamics.com) answers questions directly, and collected questions help improve the docs; 

② GitHub Issues (with dedicated bug report and feature request templates); 

③ the maintainers' emails: mason@limxdynamics.com. 

Chinese-speaking users can also join the WeChat and Feishu groups via the QR codes pinned in the README.

> Source: https://www.limxdynamics.com/en/products/fluxvla/faq