VLA GalleryAll modelsGallery

VLA Gallery / All models / New in 2026

Being-H0.5

BeingBeyond · 2026-01

VLANew in 20262Bflow matching

Trained 30 embodiments on UniHand-2.0 (16,000+ hours of human egocentric·14,000+ hours of robot data).

View architecture diagram in the galleryPaper ↗

Paper

Being-H0.5: Scaling Human-Centric Robot Learning for Cross-Embodiment GeneralizationarXiv 2601.12993

Overview

Category
New in 2026
Type
VLA
Released
2026-01
Organization
BeingBeyond
Open weights
Yes
Parameters
2B(README)/3B(HF)

Architecture

Backbone LLM
Not disclosed
Vision encoder
—
Action head
Mixture-of-Flow (unified action space) (flow matching)
Chunk
Not disclosed
Control rate
Not disclosed
Architecture (original summary)
Mixture-of-Transformers + Mixture-of-Flow; unified action space mapping heterogeneous robot control to semantically aligned slots; Universal Async Chunking

Training

Pretraining accelerators
Not disclosed (the pretraining recipe released alongside is 1,000 GPU-hours, GPU type Not disclosed)
Pretraining duration
Not disclosed
Pretraining steps & batch
Not disclosed
GPU/TPU-hours
Not disclosed
Fine-tuning
Not disclosed
Data type
Human video (egocentric) + real robot + simulation·synthetic (26%) + web vision-language
Data scale
Over 35,000 hours · 120B tokens · over 400M samples (human 16,000 hours, robot over 14,000 hours, VLM equivalent of 5,000 hours); 30 embodiments
Main datasets
UniHand-2.0, UniCraftor (200+ hours, 43 tasks)
Tasks & evaluation
LIBERO 98.9%, RoboCasa 53.9%; 5 real robots
Robot
PND Adam-U, Franka+Inspire, Unitree G1, BeingBeyond D1, LeRobot SO-101

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 used · Supervised learning (within the scope of the public paper)
RL evidence
https://arxiv.org/html/2601.12993v1
Base control
Fixed · The unified action space defines base velocity and heading slots, but experiments (including G1) control only the upper body and arms

Notes

Key features
Trained 30 embodiments on UniHand-2.0 (16,000+ hours of human egocentric·14,000+ hours of robot data). Also released a 1,000 GPU-hour pretraining recipe.
Notes
Conflicting parameters: official README 2B, HF auto label 3B, not stated in the paper (numeric column left empty). arXiv v1 2026-01-19, GitHub 2026-01-20. UniHand_Preview subset released 2026-03-20, accepted at CoRL 2026. 1,000 GPU-hours is the public recipe value; whether it matches the flagship run is unverified, so the numeric column is left empty.
Verification
Confirmed

Other models in this category