HoloBrain-0

A VLA framework Horizon Robotics open-sourced in early 2026 that feeds camera parameters and robot structure into the model as priors.

HoloBrain-0 was released as a technical report and open-sourced by the Horizon Robotics team in February 2026. Its core design explicitly feeds robot-embodiment priors into the VLA: the parameters of multiple cameras, and a URDF file describing the robot's joint structure, used to strengthen 3D spatial reasoning and adapt across different embodiments. Training follows a “pretrain then post-train” route: pretraining data comes from real robots — a bimanual Piper, AgiBot G1, and Franka — plus a bimanual UR5 and bimanual ARX in simulation, and the human-hand video dataset EgoDex, followed by post-training on specific tasks. The report says it achieves state-of-the-art results on the RoboTwin 2.0, LIBERO, and GenieSim simulation benchmarks, and also performs well on long-horizon real-robot tasks. What's open-sourced includes the pretrained model, post-training checkpoints for various simulation suites and real-robot tasks, and RoboOrchard, a full-stack toolchain covering data curation, training, and deployment.

ExampleIt has a lightweight, roughly 200-million-parameter (0.2B) version based on GroundingDINO-Tiny that performs comparably to much larger baselines with low inference latency, suited to deploying directly on a robot's onboard chip; there's also a roughly 1.1-billion-parameter version based on Qwen2.5-VL-3B.

Also called

Horizon Robotics HoloBrain, HoloBrain-0 Technical Report

Related

Vision-Language-Action Model · Cross-Embodiment · Unified Robot Description Format · RoboTwin · On-Device / Edge Deployment · Horizon Robotics

Sources

HoloBrain-0 Technical Report (arXiv 2602.12062)
HoloBrain-0 技术报告 HTML 全文 (Chinese)

As of

2026-02