VLA GalleryAll modelsGallery

VLA Gallery / All models / Humanoid VLA

Being-H0

Peking University·Renmin University of China·BeingBeyond · 2025-07

VLAHumanoid VLA1B / 8B / 14Bdiscrete tokens

Uses 'physical instruction tuning', chaining VLA pretraining on human hand videos, physical-space alignment, and robot-task post-training.

View architecture diagram in the galleryPaper ↗

Paper

Being-H0: Vision-Language-Action Pretraining from Large-Scale Human VideosarXiv 2507.15597

Overview

Category
Humanoid VLA
Type
VLA
Released
2025-07
Organization
Peking University·Renmin University of China·BeingBeyond
Open weights
Yes
Parameters
1B / 8B / 14B

Architecture

Backbone LLM
InternVL3 architecture (internal LLM not stated)
Vision encoder
InternViT-300M
Action head
MANO hand motion tokens (GRQ) (discrete tokens)
Chunk
Not disclosed
Control rate
Not disclosed
Architecture (original summary)
InternVL3 architecture (vision encoder InternViT-300M) + part-level MANO hand motion tokens (separate grouped residual quantization for wrist·fingers, codebook Kw=Kf=4096), motion at 15 FPS

Training

Pretraining accelerators
A800-80G × 32 (ViT adapter · LLM backbone trained jointly)
Pretraining duration
Not disclosed
Pretraining steps & batch
Steps not disclosed · batch 128 (AdamW, lr 1×10⁻⁵)
GPU/TPU-hours
Not disclosed
Fine-tuning
GPU and time not disclosed; post-trained with 50~100 teleoperation trajectories per task (Gello)
Data type
Motion capture / human video / real-robot teleoperation
Data scale
UniHand 440K+ trajectories·130M+ frames·1,100+ hours of video; 2.5M samples (UniHand-2.5M) used for pretraining
Main datasets
UniHand (11 sources: motion capture such as ARCTIC·FPHA·H2O·HOI4D, VR recordings EgoDex·HoloAssist, pseudo-labeled RGB video Taste-Rob)
Tasks & evaluation
Real robot Pour-Cup 100%, Close-Toolbox 85%, Pick-Place-Toy 75% (seen)/65% (unseen)/60% (clutter), Unfold-Clothes 75%, Close-Lid 60%
Robot
Franka Research 3 (7-DoF) + Inspire hand (6-DoF) + RealSense L515; not a humanoid

Inference

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

Reinforcement learning & base

Reinforcement learning
Not used · Human hand motion pretraining + imitation learning
Base control
Fixed · No base control (Franka FR3 + Inspire hand, EE pose·hand joints only)

Notes

Key features
Uses 'physical instruction tuning', chaining VLA pretraining on human hand videos, physical-space alignment, and robot-task post-training. Real-robot evaluation uses a Franka arm + Inspire hand rather than a humanoid, so it sits at the scope boundary.
Notes
Parameters (B) is the largest variant, 14B. Institutions corrected in verification by adding Peking Univ. and Renmin Univ. Sample count conflicts within the paper (§5.1 165M+ vs contribution list 150M+). Data hours 1100 is the lower bound of 'over 1,100 hours'. Weights 1B/8B/14B and 8B-Align released 2025-08-02, code MIT (README badge).
Verification
Corrections applied

Other models in this category