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

MiniVLA

Stanford(ILIAD) · 2024-12

VLAAutoregressive & latent-action VLAAbout 1Bdiscrete tokens

Shrinks OpenVLA from 7B to about 1B and adds VQ action chunks and history input.

Paper

Backbone LLM: Qwen2.5 0.5B
Vision encoder: SigLIP + DINOv2
Action head: 256-bin tokens + residual VQ chunks (discrete tokens)
Chunk: 8 (VQ chunk example)
Control rate: Not disclosed
Architecture (original summary): Qwen2.5 0.5B + same DINOv2+SigLIP ViT as OpenVLA; 256-bin extra tokens, residual VQ action chunks (e.g. horizon 8, codebook 256)

Pretraining accelerators: Not disclosed (README example uses 8 GPUs on 1 node)
Pretraining duration: Not disclosed
Pretraining steps & batch: Not disclosed
GPU/TPU-hours: Not disclosed
Fine-tuning: Not disclosed
Data type: Simulation·synthetic + real-robot teleoperation + web vision-language
Data scale: Not disclosed
Main datasets: LIBERO-90, BridgeData V2, LLaVA-1.5-Instruct VQA (base VLM)
Tasks & evaluation: LIBERO-90 82% (VQ h8 + history); SimplerEnv 4 tasks 44–70%; real Bridge V2
Robot: LIBERO·SimplerEnv simulation, Bridge V2 real-robot data

Reinforcement learning: Not used · Imitation learning (behavior cloning) only
Base control: Fixed · No base control (arm only, sim·Bridge)

Key features: Shrinks OpenVLA from 7B to about 1B and adds VQ action chunks and history input. Runs at 12.5 Hz on L40s, 2.5× faster than OpenVLA (5 Hz).
Notes: With no arXiv paper, the Stanford AI Lab blog (2024-12-12) was used as the primary source. 2-frame history input gave about 5% improvement.
Verification: Confirmed