VLA Gallery / All models / Flow·Diffusion VLA

X-VLA

Tsinghua AIR, Shanghai AI Lab, Peking University · 2025-10

VLAFlow·Diffusion VLA0.9Bflow matching

Absorbs heterogeneity across 7 platforms with per-embodiment soft prompts (0.04% of all parameters).

Paper

Backbone LLM: None (Florence-2 encoder only)
Vision encoder: Florence-Large encoder
Action head: flow matching Transformer (24 blocks) + soft prompt (flow matching)
Chunk: 30 anchors (4 s)
Control rate: Not disclosed
Architecture (original summary): Florence-Large VL encoder (encoder-only) + learnable per-embodiment soft prompts + 24-block · hidden 1024 flow matching Transformer; EE pose action = 30 anchor points of a 4-second trajectory; 10 denoising steps at inference (README example)

Pretraining accelerators: A100×64
Pretraining duration: About 4 days
Pretraining steps & batch: 200K iter, global batch 1024, bf16, lr 1e-4, wd 0.01
GPU/TPU-hours: 6,144 (Converted (approx.) · A100-hours (64 × about 4 days))
Fine-tuning: Most benchmarks batch 128·60K steps, Fractal 256·50K, Soft-Fold 256·400K (hardware·time Not disclosed); with PEFT LoRA 9M, LIBERO 93%·Simpler-WidowX 54%
Data type: Real-robot teleoperation
Data scale: 290K episodes (DROID · RoboMIND · AgiBot), 7 platforms; Soft-Fold 1,200 episodes · 2M samples · 24 hours of collection
Main datasets: DROID, RoboMIND, AgiBot World(48.8%)
Tasks & evaluation: Simpler Google Robot VM 80.4·VA 75.7 (paper)/83.5·76.4 (README), WidowX 95.8; LIBERO 98.2/98.6/97.8/97.6, average 98.1; CALVIN ABC→D 4.43; RoboTwin-2.0 Easy 70.0·Hard 39.0; NAVSIM PDMS 87.3; Soft-Fold cloth folding 2 hours 100% success (README)
Robot: WidowX, AgileX, AIRBOT real robots; cross-embodiment 7 platforms·5 arm types (+NAVSIM driving)

Inference device & speed: Not disclosed in paper; LeRobot port default chunk 32·denoising 10 (secondary)
Inference memory: Not disclosed
Weights bf16 (calc.): 1.8 GB
Where it runs: Not disclosed

Reinforcement learning: Not used · Imitation learning (behavior cloning) only
Base control: Fixed · Robot experiments use end-effector pose only (no base); NAVSIM autonomous-driving closed-loop evaluation is separate (not a robot base)

Key features: Absorbs heterogeneity across 7 platforms with per-embodiment soft prompts (0.04% of all parameters). Tuning only 1% of parameters (9M) with LoRA reached LIBERO 93%, and it won 1st place at AgiBot World Challenge @ IROS 2025.
Notes: Google Robot scores: paper 80.4/75.7 vs README 83.5/76.4. LIBERO-Long uses 97.6, consistent with the 98.1 average. Accepted at ICLR 2026 (README). App. K lists baseline retraining compute (A100×4 about 60 h·150,000 steps, π0 LoRA A800×4 about 7 h·30,000 steps).
Verification: Confirmed