VLA GalleryAll modelsGallery

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

UniVLA (BAAI)

CASIA·BAAI·Tsinghua University·HKISI · 2025-06

VLAAutoregressive & latent-action VLA8.5Bdiscrete tokens

Treats vision·language·actions as one discrete token sequence and performs world-model post-training on robot video.

View architecture diagram in the galleryPaper ↗

Paper

Unified Vision-Language-Action ModelarXiv 2506.19850

Overview

Category
Autoregressive & latent-action VLA
Type
VLA
Released
2025-06
Organization
CASIA·BAAI·Tsinghua University·HKISI
Open weights
Yes
Parameters
8.5B (same as Emu3)

Architecture

Backbone LLM
Emu3 (8.5B)
Vision encoder
VQ image tokens
Action head
DCT/FAST action tokens (interleaved video · action) (discrete tokens)
Chunk
Not disclosed
Control rate
Not disclosed
Architecture (original summary)
Unified autoregressive model based on Emu3; VQ image tokens (spatial compression 8) + DCT/FAST action tokens (vocabulary 1024), interleaved video-action sequence

Training

Pretraining accelerators
32 × A100 (40GB)
Pretraining duration
About 4–5 days (per the 50k steps run sentence)
Pretraining steps & batch
50k steps (A100 × 32 sentence); a separate sentence gives 30K steps · batch 64; 256×256 6 frames, max sequence 6400
GPU/TPU-hours
Not disclosed
Fine-tuning
CALVIN · LIBERO batch 192 · 8k steps (A100), SimplerEnv batch 128 · 20k steps · chunk 5; GPU count and time not disclosed
Data type
Real-robot teleoperation (video) + simulation·synthetic + human video
Data scale
622K videos (robot datasets + SSv2 human videos)
Main datasets
RT-1, BridgeV2, DROID, Kuka, CALVIN, LIBERO, ManiSkill2, SSv2, etc.
Tasks & evaluation
CALVIN ABC→D 4.41 · ABCD→D 4.63; LIBERO average 95.5%; SimplerEnv-Bridge 69.8%; real ALOHA
Robot
ALOHA, CALVIN · LIBERO · SimplerEnv simulation

Inference

Inference device & speed
Not disclosed
Inference memory
Not disclosed
Weights bf16 (calc.)
17 GB
Where it runs
Not disclosed

Reinforcement learning & base

Reinforcement learning
Not used · Supervised video·action generation
Base control
Fixed · Robot experiments use fixed bimanual (ALOHA); autonomous vehicle driving evaluation is separate (not a robot base)

Notes

Key features
Treats vision·language·actions as one discrete token sequence and performs world-model post-training on robot video. Reported CALVIN ABC→D average length 4.41 and LIBERO 95.5%.
Notes
The pretraining column is the world-model post-training stage on top of Emu3 (Emu3's own training compute not disclosed). Verification found '4–5 days' is the 50k steps sentence (App A) and '30K steps·batch 64' is a separate sentence (Sec 4.2), so they were not combined. Affiliations supplemented with CASIA etc. A different model from OpenDriveLab's UniVLA.
Verification
Corrections applied

Other models in this category