VLA Gallery / All models / Autoregressive & latent-action VLA

VLA-0

NVIDIA · 2025-10

VLAAutoregressive & latent-action VLA3Bdiscrete tokens

A design that makes the VLM output actions as plain-text numbers without changing its architecture or vocabulary.

Paper

Backbone LLM: Qwen2.5-3B · Qwen2.5-VL-3B config
Vision encoder: Qwen2.5-VL ViT
Action head: Integer text actions (no action head) (discrete tokens)
Chunk: Multi-step + ensembling
Control rate: Not disclosed
Architecture (original summary): Unmodified Qwen2.5-VL-3B; outputs actions as integer text (e.g. 0–1000), multi-step prediction + cross-timestep ensembling, no vocabulary change·no action head

Pretraining accelerators: N/A (no robot pretraining)
Pretraining duration: N/A
Pretraining steps & batch: N/A
GPU/TPU-hours: Not disclosed
Fine-tuning: 8 × A100 about 32 h, 64 epochs, batch 192, lr 5e-6
Data type: Simulation·synthetic + real-robot teleoperation
Data scale: LIBERO demos; 100 real SO-100 demos
Main datasets: LIBERO
Tasks & evaluation: LIBERO average 94.7%; 4 real SO-100 tasks (+12.5%p over SmolVLA)
Robot: SO-100, LIBERO simulation

Inference device & speed: RTX 5090 desktop 4 Hz (README: 6 Hz possible with TensorRT-LLM)
Inference memory: Not disclosed
Weights bf16 (calc.): 6 GB
Where it runs: Local PC

Reinforcement learning: Not used · Imitation learning (behavior cloning) only
Base control: Fixed · No base control (SO-100 arm·LIBERO)

Key features: A design that makes the VLM output actions as plain-text numbers without changing its architecture or vocabulary. Reported LIBERO 94.7% without large-scale robot pretraining.
Notes: All robot training is a single fine-tuning stage (8× A100 × about 32 h, converted (approx.) 256 A100-hours). The base model license is the Qwen Research License.
Verification: Confirmed