VLA GalleryAll modelsGallery

VLA Gallery / All models / New in 2026

HoloBrain-0

Horizon Robotics · 2026-02

VLANew in 20261.1BOther

Injects embodiment information via multi-view camera parameters and kinematic descriptions.

View architecture diagram in the galleryPaper ↗

Paper

HoloBrain-0 Technical ReportarXiv 2602.12062

Overview

Category
New in 2026
Type
VLA
Released
2026-02
Organization
Horizon Robotics
Open weights
Yes
Parameters
1.1B(QW)/0.2B(GD)

Architecture

Backbone LLM
Qwen2.5-3B · Qwen2.5-VL-3B config (QW variant)
Vision encoder
Grounding DINO Tiny (GD variant) / Qwen2.5-VL ViT (QW variant)
Action head
Joint-Graph Attention (joint angle residuals + link pose) (Other)
Chunk
Not disclosed
Control rate
Not disclosed
Architecture (original summary)
Grounding DINO Tiny (GD variant) or Qwen2.5-VL-3B (QW variant) + Joint-Graph Attention; hybrid action space (joint angle residuals in rad + link pose displacements)

Training

Pretraining accelerators
Not disclosed
Pretraining duration
Not disclosed
Pretraining steps & batch
Batch 2048·512 (by variant), steps Not disclosed
GPU/TPU-hours
Not disclosed
Fine-tuning
Not disclosed (post-training batch 256·128)
Data type
Real-robot teleoperation + simulation·synthetic (RoboTwin 2.0, Agibot-digital) + human video (EgoDex)
Data scale
156.66M frames, 3,547.4 hours (over 3,500 hours), 531,541 trajectories, 7 embodiments
Main datasets
In-house bimanual Piper, AgiBot World, DROID, RoboTwin 2.0, Agibot-digital, EgoDex
Tasks & evaluation
Real-world average progress 88.07·success 74.81% (GD variant); claims SOTA on RoboTwin 2.0·LIBERO·LIBERO-Plus·GenieSim
Robot
7 embodiments (in-house bimanual Piper, etc.)

Inference

Inference device & speed
Asynchronous inference 5–10 FPS
Inference memory
Not disclosed
Weights bf16 (calc.)
2.2 GB
Where it runs
Not disclosed

Reinforcement learning & base

Reinforcement learning
Not used · Trained with 4 supervised losses
RL evidence
https://arxiv.org/html/2602.12062
Base control
Fixed · All real tasks explicitly fixed-base tabletop manipulation (AgiBot G1 wheeled base not controlled, upper body only)

Notes

Key features
Injects embodiment information via multi-view camera parameters and kinematic descriptions. The 0.2B variant is for edge-device deployment.
Notes
arXiv 2026-02-12. Entire ecosystem open-sourced, including pretraining·post-training checkpoints. Parameters (B) is the large variant (QW 1.1B).
Verification
Confirmed

Other models in this category