Horizon RoboticsChina2026
HoloBrain-0
An AI model: a vision-language-action model (VLA) takes camera images and an instruction and turns them into movements for a robot. This page describes the software, not a robot.
Robotsnot stated
VersionsHoloBrain-0-QWHoloBrain-0-GD
The developer has released this model in several versions. Each figure applies to the version its source names.
Key figures as stated by the developer
6 of 6 stated-
Model type C
vision-language-action -
Controls A
arms and hands -
Inputs A
cameradepthlanguagerobot state - Access B open download
- Parameters A 1,100,000,000
- Base model A Qwen2.5-VL-3B
How to read the figures
- 50HzA
- The letter in the grey square after a figure points to the source the figure came from. Pages with several sources carry several letters.
Versions and notes in the key figures (6)
Model type
Developer wording: "In this work, we introduce HoloBrain-0, a comprehensive Vision-Language-Action (VLA) framework that bridges the gap between foundation model research and reliable real-world robot deployment."
Controls
Technical report: "the model generates action chunk for the subsequent" steps. Developer: "This ensures compatibility across single/dual-arm manipulators, mobile robots, and human-captured data." Real-world setup: "We conduct real-world experiments using a dual-arm AgileX Piper robot equipped with three Intel RealSense D435 RGB-D cameras."
Inputs
Technical report: "At each time step, the system receives multi-view RGB images", "and corresponding depth maps", "along with the robot's current proprioceptive joint state"; "the model is further conditioned on natural language instructions". RGB images = camera.
Access
Developer: "we fully open-source the entire HoloBrain ecosystem"; the weights are published on Hugging Face.
Parameters
Technical report: "with HoloBrain-0-QW containing 1.1B parameters and HoloBrain-0-GD containing 0.2B". HoloBrain-0-QW is shown; HoloBrain-0-GD is a variant.
Versions
- HoloBrain-0-QW
- 1,100,000,000
- HoloBrain-0-GD
- 200,000,000
Base model
Technical report, Table 2: "GD denotes GroundingDINO Tiny, QW represents Qwen2.5-VL-3B." The shown base is HoloBrain-0-QW.
Versions
- HoloBrain-0-QW
- Qwen2.5-VL-3B
- HoloBrain-0-GD
- GroundingDINO Tiny
Full specifications
10 of 13 fields stated| Field | Value | Source and caveats |
|---|---|---|
| Input and output | ||
| Model type | vision-language-action | C Developer wording: "In this work, we introduce HoloBrain-0, a comprehensive Vision-Language-Action (VLA) framework that bridges the gap between foundation model research and reliable real-world robot deployment." |
| Inputs | camera depth language robot state | A Technical report: "At each time step, the system receives multi-view RGB images", "and corresponding depth maps", "along with the robot's current proprioceptive joint state"; "the model is further conditioned on natural language instructions". RGB images = camera. |
| Controls | arms and hands | A Technical report: "the model generates action chunk for the subsequent" steps. Developer: "This ensures compatibility across single/dual-arm manipulators, mobile robots, and human-captured data." Real-world setup: "We conduct real-world experiments using a dual-arm AgileX Piper robot equipped with three Intel RealSense D435 RGB-D cameras." |
| Size and speed | ||
| Parameters | 1,100,000,000 | A Technical report: "with HoloBrain-0-QW containing 1.1B parameters and HoloBrain-0-GD containing 0.2B". HoloBrain-0-QW is shown; HoloBrain-0-GD is a variant. Versions
|
| Where it runs and what it is built on | ||
| On robot or cloud | not stated | Note Technical report: the action expert gives "a lightweight architecture that enables efficient deployment on resource-constrained edge devices"; the abstract says of the 0.2B variant (HoloBrain-0-GD): "enabling low-latency on-device deployment". A |
| Base model | Qwen2.5-VL-3B | A Technical report, Table 2: "GD denotes GroundingDINO Tiny, QW represents Qwen2.5-VL-3B." The shown base is HoloBrain-0-QW. Versions
|
| Openness | ||
| Access | open download | B Developer: "we fully open-source the entire HoloBrain ecosystem"; the weights are published on Hugging Face. |
| License | Apache-2.0 | B Model cards: "license: apache-2.0"; the code licence file reads "Apache License". |
| Free or paid | fully open source | C Developer: "we fully open-source the entire HoloBrain ecosystem"; both model cards give "license: apache-2.0". |
| The developer’s own claims | ||
| Learns new tasks | needs retraining | A Developer claim: post-training per task, "We collect only 200 episodes per task and train a single multi-task model, aiming to validate the model's learning efficacy with limited post-training data." |
| Works across robots | yes | A Developer claim: "our model effectively unifies training across heterogeneous robots". The model card gives deployment commands for two arms: "Choose the command that matches your robot arm:" "Piper robot arm" and "PiperX robot arm". Yes only if the developer itself says the model runs on different robot models, for example robots from two manufacturers or two of one manufacturer’s models. Several units of the same robot do not count. |
Not stated: Degrees of freedom controlled, Control rate
Notes
- Horizon Robotics describes HoloBrain-0 as a Vision-Language-Action framework for real-world robot deployment.
- The HoloBrain-0 technical report was submitted to arXiv on 12 February 2026.
- Horizon Robotics names two sizes: HoloBrain-0-QW is built on Qwen2.5-VL-3B and HoloBrain-0-GD on GroundingDINO Tiny.
- The Horizon Robotics website footer gives the registered company name, Beijing Horizon Information Technology Co., Ltd.
Sources
-
A
arxiv.
org/ · Retrievedhtml/ 2602. 12062 -
B
huggingface.
co/ · RetrievedHorizonRobotics/ HoloBrain_ v0. 0_ Qwen -
C
arxiv.
org/ · Retrievedabs/ 2602. 12062
One line per source, not per figure: a source has one retrieval date. The letter beside a figure points to the line it came from.