What Is FLOWER VLA?

FLOWER is introduced in arXiv 2509.04996 by Moritz Reuss, Hongyi Zhou, Marcel Rühle, Ömer Erdinç Yağmurlu, Fabian Otto and Rudolf Lioutikov (v1 5 Sep 2025), published at CoRL 2025 per the arXiv comments. The project page credits the Intuitive Robots Lab at the Karlsruhe Institute of Technology and Microsoft Research. Read: the arXiv abstract page, project page and repository README; the full paper text was not read.

Third-party research; Firsthand is not affiliated and claims no use, training or compatibility with it. Firsthand captures first-person human video, not robot data, fast, on demand and to a written spec: 21 keypoints per hand with a visibility flag, wrist trajectories and action segments, plus one task instruction per episode.

Efficiency. The abstract says current diffusion-based VLA policies need multi-billion-parameter models and massive datasets, and proposes two contributions to cut that cost: intermediate-modality fusion and action-specific Global-AdaLN conditioning. The title frames the goal as democratizing generalist robot policies.

Intermediate fusion prunes layers of a pretrained language model, up to 50% per the abstract and 30 to 50% of the final layers per the project page, and conditions the action head on intermediate embeddings instead. Global-AdaLN shares modulation weights, which the abstract says cuts parameters by 20%. The project page adds lightweight LoRA adapters and separate encoders and decoders for different action spaces.

The README describes a Florence-2-large-based vision-language model with a rectified flow action head. Listed parts are a 350M-parameter DaViT image encoder, language conditioning through cross-attention, half of the Florence-2 language layers at 205M, and action chunking. It says all CALVIN and LIBERO results use four denoising steps and that inference needs under 3GB of GPU memory with a single image.

The abstract states pretraining took 200 H100 GPU hours. The project page says matching its fine-tuned results needs about four hours on four GPUs. The README repository covers fine-tuning on CALVIN and LIBERO, and points to a separate codebase for pretraining and Aloha fine-tuning and to Hugging Face checkpoints. The model cards were not opened.

The abstract claims competitive performance with larger VLAs across 190 tasks spanning ten simulation and real-world benchmarks, and a new state of the art of 4.53 on CALVIN ABC. The project page adds 20 real-world tasks, three long-horizon sequences and a 28% average gain over OpenVLA in its generalization scenarios, and labels some of its own videos as failures. The README’s CALVIN table lists a different ABC figure, so none is treated as settled. These are the authors’ numbers.

The site’s entries describe pi0 as adding a flow-matching action expert to a pretrained vision-language model, and SmolVLA as a compact base model on SmolVLM2-500M-Video-Instruct. FLOWER takes a different route, pruning a Florence-2 backbone and putting its emphasis on parameter and compute efficiency. The table sets them side by side; this page ranks none of them.

The sources read call it a flow policy with a rectified flow action head. The site’s flow-matching entry covers that family. The exact training objective sits in the full paper, which was not read, so it is Not stated here.

The pretraining data composition, the full paper’s method detail and the Hugging Face model cards were not read. FLOWER’s stated inputs are robot trajectories and images with language instructions, so the sources do not say whether first-person human video would suit it. The README’s MIT licence applies to the repository; terms for the datasets it names are Not stated here.

FLOWER’s stated inputs are robot data, so human video is not its stated input. Firsthand captures human first-person video, not robot data: hand pose, wrist trajectories, action segments and one task instruction per episode; robot actions are Not stated. Write tasks, environments and signals into a spec and confirm on a Spec Pilot. Median time to first validated batch: 48 hours where coverage is deep, three to four weeks where thin. Formats include RLDS; matching FLOWER’s expected fields is the buyer’s to check.

What is FLOWER VLA?

A 950M-parameter vision-language-action flow policy (arXiv 2509.04996, CoRL 2025) that prunes a pretrained vision-language model and conditions a flow-based action head on intermediate embeddings.

What makes FLOWER efficient?

Per the abstract, intermediate-modality fusion prunes up to 50% of language-model layers and Global-AdaLN cuts parameters by 20%, with pretraining in 200 H100 GPU hours.

What does FLOWER claim on results?

The abstract states competitive performance across 190 tasks on ten benchmarks and a new state of the art on CALVIN ABC. These are the authors’ figures, and the README lists a different ABC number, so none is reproduced as settled.

Is FLOWER open source?

The official repository states the MIT licence and links Hugging Face checkpoints from its README. The model cards were not opened here.

Does Firsthand use FLOWER, and can I request a custom dataset?

Firsthand is not affiliated and claims no use, training or compatibility. You can request a custom dataset, confirmed first on a Spec Pilot.