VLA Gallery / All models / Flow·Diffusion VLA
EO-1
EO-1 (EmbodiedOneVision)
EO-Robotics† (README contact pjlab.org.cn, mirror SHAILAB-IPEC) · 2025-08
VLAFlow·Diffusion VLA3Bhybrid
Jointly learns reasoning and actions in one decoder from interleaved sequences mixing vision·text·actions.
Paper
EmbodiedOneVision: Interleaved Vision-Text-Action Pretraining for General Robot ControlarXiv 2508.21112
Overview
- Category
- Flow·Diffusion VLA
- Type
- VLA
- Released
- 2025-08
- Organization
- EO-Robotics† (README contact pjlab.org.cn, mirror SHAILAB-IPEC)
- Open weights
- Yes
- Parameters
- 3B
Architecture
- Backbone LLM
- Qwen2.5-3B · Qwen2.5-VL initialization
- Vision encoder
- Qwen2.5-VL ViT
- Action head
- Autoregressive tokens + flow matching (10 steps) (hybrid)
- Chunk
- 16 steps
- Control rate
- Not disclosed
- Architecture (original summary)
- Single decoder-only Transformer initialized from Qwen2.5-VL; discrete autoregressive decoding + continuous flow matching denoising, interleaved vision-text-action sequences; chunk 16, 10 denoising steps
Training
- Pretraining accelerators
- '128+ GPUs' (README, GPU model not disclosed; flash-attn 3 recommends H100/H800)
- Pretraining duration
- Not disclosed
- Pretraining steps & batch
- 5 epochs, variable-length packing (average 16,384), batch 1, lr 5×10⁻⁵ (LM · projector)/1×10⁻⁵ (vision)
- GPU/TPU-hours
- Not disclosed
- Fine-tuning
- LIBERO: GPU×8, 64 per device (effective 512), 50 epochs; real robot 150 demos per task·25 epochs (subtask:full 9:1); time·GPU model not disclosed
- Data type
- Real-robot teleoperation + web vision-language + interleaved embodied data
- Data scale
- 135B tokens total: robot 1.2M episodes (127.3B tokens), web 5.7M samples (7.1B), EO-Data1.5M (1.0B)
- Main datasets
- AgiBotWorld, OXE, RoboMIND, SO100-Community, IPEC-Franka; LLaVA-1.5, LLaVA-Video-178K, PixMo-Points, RefCOCO, RoboVQA; EO-Data1.5M
- Tasks & evaluation
- LIBERO 98.2±0.3% (repo 50 epochs 99.7/99.8/99.2/94.8); SimplerEnv WidowX 72.7%, Google VM 76.5%·VA 63.0%; RoboVQA BLEU-4 58.5, ERQA 45.5; EO-Bench Spatial 36.4·Temporal 38.9; real robot 86.0%
- Robot
- Franka Panda, WidowX 250 S, AgiBot G-1, LeRobot SO100
Inference
- Inference device & speed
- Single RTX 4090 6 GB (README about 6.5 GB); speed not disclosed
- Inference memory
- 6 GB (README about 6.5 GB, RTX 4090)
- Weights bf16 (calc.)
- 6 GB
- Where it runs
- Not disclosed
Reinforcement learning & base
- Reinforcement learning
- Not used · Imitation learning (behavior cloning) only
- Base control
- Fixed · No base control (manipulation tasks on 4 platforms including AgiBot G-1)
Notes
- Key features
- Jointly learns reasoning and actions in one decoder from interleaved sequences mixing vision·text·actions. Trained on 135B tokens (1.2M robot episodes), reaching LIBERO 98.2%.
- Notes
- No training-GPU statement in paper ('128+ GPUs' is only in README). Institution names could not be confirmed from paper extraction†. Revised up to arXiv v5 (2026-02-25).
- Verification
- Confirmed