VLA Gallery / All models / Flow·Diffusion VLA
F1-VLA
Shanghai AI Laboratory, Harbin Institute of Technology (Shenzhen) · 2025-09
VLAFlow·Diffusion VLA4.2Bflow matching
A VLA that generates future visual scenes (foresight) in a next-scale manner before actions and uses them for planning.
Paper
F1: A Vision-Language-Action Model Bridging Understanding and Generation to ActionsarXiv 2509.06951
Overview
- Category
- Flow·Diffusion VLA
- Type
- VLA
- Released
- 2025-09
- Organization
- Shanghai AI Laboratory, Harbin Institute of Technology (Shenzhen)
- Open weights
- Yes
- Parameters
- 4.2B
Architecture
- Backbone LLM
- Gemma · PaliGemma initialization
- Vision encoder
- SigLIP (PaliGemma)
- Action head
- foresight generation + flow matching expert (flow matching)
- Chunk
- Not disclosed
- Control rate
- Not disclosed
- Architecture (original summary)
- Mixture-of-Transformer: understanding expert (PaliGemma initialization) + foresight generation expert (VAR residual VQ-VAE next-scale, 4 scales at inference) + flow matching action expert (inherited from π0)
Training
- Pretraining accelerators
- Not disclosed (no GPU type·count·time in App. B)
- Pretraining duration
- Not disclosed
- Pretraining steps & batch
- Stage I 512K steps·batch 1280·lr 3.0e-4 (cosine); Stage II 100K steps·batch 2880·lr 5.0e-5 (constant)
- GPU/TPU-hours
- Not disclosed
- Fine-tuning
- Stage III batch 128·lr 5.0e-5 (cosine), hardware·time not disclosed; post-training on about 1.8K in-house trajectories
- Data type
- Real-robot teleoperation + simulation (LIBERO)
- Data scale
- 330.9K trajectories · 73.8M frames, 136 tasks, 5 embodiments (AgiBot-World 187K/66.4M, Bridge-v2 53.2K/1.9M, Fractal 87.2K/3.8M, LIBERO 1.7K/0.3M)
- Main datasets
- AgiBot World, OXE Bridge v2, OXE Fractal, LIBERO
- Tasks & evaluation
- LIBERO 98.2/97.8/95.4/91.3, average 95.7; SimplerEnv Bridge 72.9; Genie real-robot 9 tasks average 82.2% (grasp 92.6%)
- Robot
- Genie G1, Franka, ARX LIFT II, WidowX, Google Robot (single-arm·bimanual)
Inference
- Inference device & speed
- Not disclosed (limits foresight to 4 scales for real-time control)
- Inference memory
- Not disclosed
- Weights bf16 (calc.)
- 8.4 GB
- Where it runs
- Not disclosed
Reinforcement learning & base
- Reinforcement learning
- Not used · Imitation learning (behavior cloning) only
- Base control
- Fixed · No mention of base movement (Genie-1·ARX LIFT II manipulation; dynamic tasks use a conveyor belt)
Notes
- Key features
- A VLA that generates future visual scenes (foresight) in a next-scale manner before actions and uses them for planning. Achieved 82.2% average success rate on 9 real Genie tasks.
- Notes
- Initialized from the π0 action expert·PaliGemma (README requires lerobot/pi0_base). App. B hyperparameters newly confirmed in verification (Unverified in notes).
- Verification
- Corrections applied