VLA GalleryAll modelsGallery

VLA Gallery / All models / New in 2026

UnifoLM-VLA-0

Unitree · 2026-01

VLANew in 2026Not disclosedNot disclosed

Performs 12 categories of manipulation tasks on Unitree G1 with a single policy.

View architecture diagram in the galleryGitHub ↗

Paper

No paper

Overview

Category
New in 2026
Type
VLA
Released
2026-01
Organization
Unitree
Open weights
Yes
Parameters
Not disclosed

Architecture

Backbone LLM
Qwen2.5-VL family (inherited via code, unclear)
Vision encoder
—
Action head
Chunked action prediction (Not disclosed)
Chunk
Not disclosed
Control rate
Not disclosed
Architecture (original summary)
VLM-based (code inherits Qwen2.5-VL, exact base unclear) + chunked action prediction (NUM_ACTIONS_CHUNK)

Training

Pretraining accelerators
Not disclosed
Pretraining duration
Not disclosed
Pretraining steps & batch
Not disclosed
GPU/TPU-hours
Not disclosed
Fine-tuning
Not disclosed
Data type
Real robot (G1 manipulation data, collection method not stated) + dynamics prediction data
Data scale
Not disclosed
Main datasets
Not disclosed
Tasks & evaluation
Single G1 policy across 12 categories (block stacking · towel folding · table wiping · two-robot cooperative cleaning, etc.); LIBERO evaluated but no figures in README
Robot
Unitree G1

Inference

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

Reinforcement learning & base

Reinforcement learning
Not disclosed · Training method not disclosed
Base control
Fixed · No locomotion control (12 G1 upper-body tabletop manipulation tasks)

Notes

Key features
Performs 12 categories of manipulation tasks on Unitree G1 with a single policy. Continued pretraining of the VLM on robot manipulation data and full-dynamics prediction data.
Notes
Training and inference code and weights released 2026-01-29. Checkpoints UnifoLM-VLM-Base, UnifoLM-VLA-Base, UnifoLM-VLA-LIBERO. License CC BY-NC-SA 4.0.
Verification
Confirmed

Other models in this category