VLA Gallery / All models / Flow·Diffusion VLA
InternVLA-M1
Shanghai AI Laboratory (Intern Robotics) · 2025-10
VLAFlow·Diffusion VLAAbout 4.1Bdiffusion
A dual system that first pretrains the VLM on spatial grounding to locate key objects, then attaches a DiT action expert.
Paper
InternVLA-M1: A Spatially Guided Vision-Language-Action Framework for Generalist Robot PolicyarXiv 2510.13778
Overview
- Category
- Flow·Diffusion VLA
- Type
- VLA
- Released
- 2025-10
- Organization
- Shanghai AI Laboratory (Intern Robotics)
- Open weights
- Yes
- Parameters
- About 4.1B; diffusion policy 86M, DINOv2 21M, state encoder 0.4M, querying transformer 8.7 MB (unit as in source)
Architecture
- Backbone LLM
- Qwen2.5-3B · Qwen2.5-VL-3B config
- Vision encoder
- Qwen2.5-VL ViT + DINOv2 (21M)
- Action head
- DiT diffusion expert 86M (DDIM 10) (diffusion)
- Chunk
- 16 · 8 (LIBERO)
- Control rate
- Not disclosed
- Architecture (original summary)
- Qwen2.5-VL-3B-Instruct (System 2, spatial prompting) → querying transformer → DiT diffusion policy action expert (System 1, 86M; DINOv2 21M·state encoder 0.4M); chunk 16 (SimplerEnv·real robot)/8 (LIBERO), DDIM 10 steps (README)
Training
- Pretraining accelerators
- Stage 1 not disclosed; Stage 2 (action post-training) A100×16
- Pretraining duration
- Not disclosed
- Pretraining steps & batch
- Stage 2: 50k steps (about 2.5 epochs), batch robot 256 + multimodal 64
- GPU/TPU-hours
- Not disclosed
- Fine-tuning
- LIBERO A100×8·batch 128·about 30K steps·about 20 h; large-scale pick-place·real robot A100×16·batch 256·20K steps; README: verified on A100·RTX 4090
- Data type
- Web vision-language (spatial grounding) + simulation·synthetic + real-robot teleoperation
- Data scale
- Multimodal 3M+ (spatial grounding 2.3M + multimodal understanding 0.7M); synthetic InternData-M1 244K closed-loop samples (14K objects, 211 tables, 1.6K textures, 87 lighting); OXE; real teleoperation 6 h + long-horizon tasks 22 h (about 500 demonstrations per task)
- Main datasets
- RefCOCO, RoboRef, A0 ManiSkill, MolmoAct, Pixmo-Points; InternData-M1; OXE(fractal_rt_1, bridge)
- Tasks & evaluation
- SimplerEnv Google VM 80.7·VA 76.0, WidowX 71.7; LIBERO 95.9; synthetic large-scale pick-place; Franka FR3 real-robot cluttered pick-place (unseen objects·new layouts +20.6%), long-horizon tasks (tidying·drawer·sandwich)
- Robot
- Google Robot, WidowX, Franka FR3 (single arm)
Inference
- Inference device & speed
- Single RTX 4090 about 12 GB; VLM about 10 FPS with FlashAttention
- Inference memory
- About 12 GB (RTX 4090)
- Weights bf16 (calc.)
- 8.2 GB
- Where it runs
- Not disclosed
Reinforcement learning & base
- Reinforcement learning
- Not used · Imitation learning (behavior cloning) only
- Base control
- Fixed · No base control (single arm)
Notes
- Key features
- A dual system that first pretrains the VLM on spatial grounding to locate key objects, then attaches a DiT action expert. Improved real-robot pick-place on unseen objects by 20.6% with 244K synthetic samples.
- Notes
- The detailed QA counts (about 637K·879K·684K·832K) do not sum to 2.3M (they appear to span the whole 3M+ bundle). Stage 1·synthetic data generation compute not disclosed. Weights released per README '0918' (year not given), so the release month follows arXiv v1 (2025-10-15). Data hours exist only for the real-robot part, so null.
- Verification
- Confirmed