VLA Gallery / All models / Flow·Diffusion VLA

FLOWER

KIT Intuitive Robots Lab (1 co-author from Microsoft Research) · 2025-09

VLAFlow·Diffusion VLA947Mflow matching

Pretrained in about 200 H100 GPU-hours, cutting parameters by over 20% via VLM mid-layer fusion and per-action-space Global-AdaLN.

Paper

Category: Flow·Diffusion VLA
Type: VLA
Released: 2025-09
Organization: KIT Intuitive Robots Lab (1 co-author from Microsoft Research)
Open weights: Yes
Parameters: 947M (body)/950M (abstract); ViT 360M (paper)/350M (README), VLM 205M, Flow Transformer 339M, Global-AdaLN 28.3M

Backbone LLM: None (Florence-2 VLM 205M, decoder removed)
Vision encoder: Florence-2-L ViT (360M)
Action head: Flow Transformer 339M (4–8 steps) (flow matching)
Chunk: 20 steps
Control rate: 3–50 Hz
Architecture (original summary): Florence-2-L with decoder removed (ViT 360M + VLM 205M, intermediate-layer fusion) + 18-layer·1024-dim Flow Transformer (339M), Action-Space Global-AdaLN-Zero; rectified flow, chunk 20, denoising 4 (single arm)/8 (high-frequency bimanual), 3~50 Hz

Pretraining accelerators: H100×4 (1 node)
Pretraining duration: 48 h
Pretraining steps & batch: 360,000 steps (300k~400k depending on data mix), batch 256 × grad accum 4 = 1024, lr FlowT 1e-4 · VLM 1e-5
GPU/TPU-hours: 200 (Stated · H100-hours (about 200; 4 × 48 h))
Fine-tuning: CALVIN GPU×4 · batch 8 per GPU · 40k steps; 30k per LIBERO suite (LIBERO-90 50k); real kitchen 50,000 steps; time not disclosed; repo training needs 24GB+ VRAM
Data type: Real-robot teleoperation
Data scale: 8 public datasets about 250k trajectories (delta-EEF 74%, single-arm joints 26%); real robot 417 trajectories (45 min of human demonstrations)
Main datasets: OXE soup: Bridge 28.62%, Fractal 24.68%, DROID 23.50%, etc.
Tasks & evaluation: CALVIN ABCD→D 4.67, ABC→D 4.53 (abstract)/4.54 (README), D→D 4.36; LIBERO-10 94.5% (README)/94.9% (paper), LIBERO-90 93.4%, Spatial 97.2%, Object 99.3%, Goal 96.9%; SIMPLER Bridge 40.0%, Google Robot 31.9%; Aloha sim insertion 0.82; real kitchen average 61%
Robot: Single arm (Franka Panda real kitchen), bimanual (Aloha sim), 4 embodiments

Inference device & speed: RTX 4090: 311.04 Hz, 0.052 s, 1,848 MB (27.6% of π0, 12.7% of OpenVLA)
Inference memory: 1,848 MB (RTX 4090)
Weights bf16 (calc.): 1.9 GB
Where it runs: Local PC

Reinforcement learning: Not used · Imitation learning (behavior cloning) only
Base control: Fixed · No base control (Franka arm·bimanual sim)

Key features: Pretrained in about 200 H100 GPU-hours, cutting parameters by over 20% via VLM mid-layer fusion and per-action-space Global-AdaLN. Runs at 311 Hz on RTX 4090 using only 27.6% of π0's memory.
Notes: Parameters (B) is the in-text 947M. Sources differ: 947M/950M, ViT 360M/350M, LIBERO-Long 94.9/94.5, CALVIN ABC 4.53/4.54. arXiv v1 dateline unverified (HF Papers 2025-09-05).
Verification: Confirmed