From-scratch pretrain backbones for a FLOWER-style Vision-Language-Action (VLA) policy, released alongside a master's thesis comparing a multi-step Rectified Flow (RF) action head against a single-step Improved MeanFlow (iMF) action head. Both backbones share the FLOWER architecture (Florence-2-large VLM with the text decoder removed, feeding an 18-block causally masked DiT action expert, hidden size 1024) and differ only in the action-head objective.

Both variants were pretrained from scratch on the same Open-X Embodiment "flowereef" mixture (nine public RLDS datasets, e.g. bridge, fractal, bc_z, eef_droid, dobbe, stanford_hydra, cmu_play_fusion, libero_10, libero_goal), so RF and iMF see an identical realised data mix. They are intended as starting points for downstream fine-tuning (e.g. LIBERO, CALVIN), not as ready-to-deploy task policies.

Each subfolder is self-contained: weights-only model.safetensors plus the resolved Hydra config.yaml used at pretraining time.

A convenience downloader is shipped in the code repository:

Pretrained on 4xA100 GPUs (Leonardo HPC) in bfloat16. The RF arm follows the FLOWER recipe; the iMF arm replaces the velocity head with the Improved MeanFlow objective. Released under MIT, consistent with the upstream FLOWER codebase. See the thesis for the full pretraining mixture, training recipe, and downstream evaluation.