VLA Gallery / All models / Flow·Diffusion VLA
WALL-OSS
X Square Robot · 2025-09
VLAFlow·Diffusion VLA4Bhybrid
Uses Uni-CoT, which produces reasoning (CoT), subtask plans, and continuous actions in one stream, with separate FFNs for actions and VL.
Paper
Igniting VLMs toward the Embodied SpacearXiv 2509.11766
Overview
- Category
- Flow·Diffusion VLA
- Type
- VLA
- Released
- 2025-09
- Organization
- X Square Robot
- Open weights
- Yes
- Parameters
- 4B (HF card); backbone Qwen2.5-VL-3B
Architecture
- Backbone LLM
- Qwen2.5-3B · Qwen2.5-VL-3B config
- Vision encoder
- Qwen2.5-VL ViT
- Action head
- FAST tokens + flow matching (MoE with separate FFNs) (hybrid)
- Chunk
- Not disclosed
- Control rate
- Not disclosed
- Architecture (original summary)
- Qwen2.5-VL-3B + MoE-style split Action FFN·VL FFN; Uni-CoT (instruction→reasoning→subtask plan→continuous actions); FAST discrete tokens in the Inspiration stage, flow matching continuous actions in the Integration stage (stage 1 VLM frozen·flow head only, stage 2 joint); chunk·frequency not disclosed
Training
- Pretraining accelerators
- Not disclosed
- Pretraining duration
- Not disclosed
- Pretraining steps & batch
- Not disclosed
- GPU/TPU-hours
- Not disclosed
- Fine-tuning
- Not disclosed; episodes per task Place-by-Color 500, Block-Spell 1600, Set-Table 1500, Tidy-Bedroom 1000, Collect-Waste 900, Pick-Place-Cup 500; LeRobot guide peak VRAM about 24~40 GB, A100 40GB+ (secondary)
- Data type
- Real-robot teleoperation + web vision-language
- Data scale
- Over 10,000 h (Sec. 4)/'tens of thousands of hours' (Sec. 4.4)
- Main datasets
- In-house collection + AgiBotWorld, DROID, BC-Z, RH20T, Furniture_bench, Fractal, Bridge v2; CapsFusion, Cambrian, PixMo, RoboPoint, Robo2VLM, COCO, VQAv2
- Tasks & evaluation
- Embodied VQA: Object Grounding 91.6%, Scene Captioning 87.6%, Action Planning 69.0% (base 46.1/57.7/59.8); 6 real tasks (Set-Table · Tidy-Bedroom, etc.) task progress seen 85% · novel 61%
- Robot
- Desktop arm, mobile stand, wheeled bimanual, wheeled humanoid
Inference
- Inference device & speed
- Not disclosed
- Inference memory
- Not disclosed
- Weights bf16 (calc.)
- 8 GB
- Where it runs
- Not disclosed
Reinforcement learning & base
- Reinforcement learning
- Not used · Joint supervised learning of reasoning, FAST, and flow
- Base control
- Wheeled (Direct base velocity/pose output) · In the Collect-waste task the model moves to the trash bin; base command format not stated in the paper (DoF template masking)
Notes
- Key features
- Uses Uni-CoT, which produces reasoning (CoT), subtask plans, and continuous actions in one stream, with separate FFNs for actions and VL. Reports seen-task progress of 85% on 6 real-world tasks, higher than π0 and Diffusion Policy.
- Notes
- Data scale conflicts within the paper between over 10,000 h and tens of thousands of hours (checked in original). Training compute Not disclosed in both paper·README. Weights wall-oss-flow, wall-oss-fast.
- Verification
- Confirmed