VLA Gallery / All models / Flow·Diffusion VLA
DiVLA (Diffusion-VLA)
Midea Group, East China Normal University, Shanghai University · 2024-12
VLAFlow·Diffusion VLADiVLA-2B / 7B / 72B variantsdiffusion
Injects VLM-generated reasoning sentences into the diffusion policy head via FiLM, with training data converted into a reasoning-inclusive format by GPT-4o.
Paper
Diffusion-VLA: Scaling Robot Foundation Models via Unified Diffusion and AutoregressionarXiv 2412.03293
Overview
- Category
- Flow·Diffusion VLA
- Type
- VLA
- Released
- 2024-12
- Organization
- Midea Group, East China Normal University, Shanghai University
- Open weights
- Unverified
- Parameters
- DiVLA-2B / 7B / 72B variants (the body's Qwen2-VL description says '2B, 8B, 72B'); head size not disclosed
Architecture
- Backbone LLM
- Qwen2 1.5B / 7B / 72B · Qwen2-VL-2B/7B/72B config
- Vision encoder
- Qwen2-VL ViT
- Action head
- Diffusion Policy head + FiLM reasoning injection (diffusion)
- Chunk
- Not disclosed
- Control rate
- Not disclosed
- Architecture (original summary)
- Qwen2-VL backbone + randomly initialized Diffusion Policy head (Chi et al. 2023 design) + FiLM reasoning injection module; chunk length not disclosed
Training
- Pretraining accelerators
- Not disclosed
- Pretraining duration
- Not disclosed
- Pretraining steps & batch
- Not disclosed
- GPU/TPU-hours
- Not disclosed
- Fine-tuning
- GPU and time not disclosed; fixed lr 2e-5, 20 epochs (which of DiVLA · OpenVLA it applies to is ambiguous in the original too)
- Data type
- Real-robot teleoperation + GPT-4o-generated reasoning annotations
- Data scale
- Pretraining DROID (2B·7B), OXE+DROID (72B), scale not stated; fine-tuning sorting 500, multitask 580, bussing 400 trajectories
- Main datasets
- DROID, Open X-Embodiment
- Tasks & evaluation
- Franka factory sorting overall average 66.2%; zero-shot bin picking of 102 unseen objects 63.7%; AgileX bimanual table bussing seen 72.9%
- Robot
- Franka (single arm), AgileX (bimanual)
Inference
- Inference device & speed
- 1 × A6000: DiVLA-2B 82 Hz, 7B 42 Hz (74/30 Hz without vLLM); 6× faster than OpenVLA
- Inference memory
- Not disclosed
- Weights bf16 (calc.)
- 144 GB
- Where it runs
- Not disclosed
Reinforcement learning & base
- Reinforcement learning
- Not used · Imitation learning (behavior cloning) only
- Base control
- Fixed · No base control (Franka·AgileX bimanual manipulation)
Notes
- Key features
- Injects VLM-generated reasoning sentences into the diffusion policy head via FiLM, with training data converted into a reasoning-inclusive format by GPT-4o. Achieved 63.7% on zero-shot bin picking of 102 unseen objects.
- Notes
- Parameters (B) based on the name of the largest variant, DiVLA-72B (unclear whether the head is included). The text's '2B, 8B, 72B' describes Qwen2-VL sizes and differs from the variant names (inconsistency in the original). DiVLA training scripts added to the DexVLA repository (2025-03-11).
- Verification
- Confirmed