VLA Gallery / All models / Flow·Diffusion VLA
GraspVLA
Galbot, Peking University, HKU, BAAI · 2025-05
VLAFlow·Diffusion VLALLM 1.8Bflow matching
Without real-robot data, co-trains synthetic SynGrasp-1B (10M trajectories) and web grounding (GRIT) for zero-shot grasping.
Paper
GraspVLA: a Grasping Foundation Model Pre-trained on Billion-scale Synthetic Action DataarXiv 2505.03233
Overview
- Category
- Flow·Diffusion VLA
- Type
- VLA
- Released
- 2025-05
- Organization
- Galbot, Peking University, HKU, BAAI
- Open weights
- Yes
- Parameters
- LLM 1.8B; total·action expert size not disclosed
Architecture
- Backbone LLM
- InternLM2 1.8B
- Vision encoder
- DINOv2 + SigLIP (frozen)
- Action head
- bbox → grasp pose → flow chunk (flow matching)
- Chunk
- 4 steps
- Control rate
- 10 Hz (data)
- Architecture (original summary)
- InternLM2 1.8B (trained) + frozen DINOv2·SigLIP; Progressive Action Generation: 2D bbox → grasp pose → flow matching action chunk (4 steps); actions stored at 10 Hz
Training
- Pretraining accelerators
- Model training Not disclosed (data generation on RTX 4090×160)
- Pretraining duration
- Model training Not disclosed (data generation 10 days)
- Pretraining steps & batch
- Batch 384, lr 1.6e-4, success stabilizes after 120k steps; README example 200K steps·8 GPUs/node (not a paper value)
- GPU/TPU-hours
- Not disclosed
- Fine-tuning
- Compute not disclosed; post-training tasks 1–2: 100 demonstrations, task 3: 10 per bottle; 5k synthetic trajectories for new arm·camera adaptation (1 day on 1 RTX 4090)
- Data type
- Simulation·synthetic + web vision-language
- Data scale
- SynGrasp-1B 10M trajectories (about 100 frames per trajectory, about 1B frames), 10,680 object meshes·240 categories; GRIT (mixing ratio not disclosed)
- Main datasets
- SynGrasp-1B, GRIT
- Tasks & evaluation
- Real-robot zero-shot grasping: synthetic categories 93.3%, web categories 93.3%; LIBERO Long 82.0·Goal 91.2·Object 94.1
- Robot
- Franka Panda (single arm), adapted to UR5e
Inference
- Inference device & speed
- L40s + Torch Compile about 200 ms (vision 9·bbox 72·grasp pose 50·flow 64 ms), about 9 GB; 5 Hz (AnyGrasp 37 Hz)
- Inference memory
- About 9 GB (L40s)
- Weights bf16 (calc.)
- Not disclosed
- Where it runs
- Not disclosed
Reinforcement learning & base
- Reinforcement learning
- Not used · Imitation learning on synthetic data
- Base control
- Fixed · No base control (Franka·UR5e arm grasping)
Notes
- Key features
- Without real-robot data, co-trains synthetic SynGrasp-1B (10M trajectories) and web grounding (GRIT) for zero-shot grasping. Data generation took 10 days on 160 RTX 4090s, about $5,000.
- Notes
- Do not confuse data-generation compute with model-training compute. Zero-shot grasping 93.3% is an extracted Table 1 value; cross-checking against the original is recommended. Weights cc-by-nc-4.0 (released 2025-07-25).
- Verification
- Confirmed