Introduction
PhysBrain 1.5 is an embodied vision-language model that unifies three core capabilities essential for physical interaction: understanding the observed world, generating goal-directed actions, and predicting how the environment will evolve. Its design mirrors the closed physical loop of agent-environment interaction: observations guide reasoning and action, actions alter the environment, and the updated state feeds back as the next observation.
To implement this loop in a single model, PhysBrain 1.5 extends pretrained vision-language models with dedicated action and visual-state tokens. Language responses, structured spatial outputs, end-effector trajectories, and future world states are all formulated as discrete tokens. These modalities are jointly learned using a single shared autoregressive backbone under a unified next-token prediction objective, with no task-specific heads.
Try PhysBrain 1.5
Experience PhysBrain 1.5 directly in your browser with 🤗 PhysBrain 1.5-8B Demo. Upload an image and try visual-spatial understanding, embodied reasoning, and planning without setting up a local environment.
Evaluation
Embodied behavior can be viewed as a recurrent physical loop: the agent perceives a 2D scene, constructs a 3D spatial representation, plans an action strategy, localizes the target point, generates a motion trajectory, executes the action, and observes the resulting state transition. The benchmark results and qualitative examples below evaluate PhysBrain 1.5's capabilities across this loop in seven categories: five in embodied understanding, plus action prediction and future-state prediction.
Embodied Understanding
We evaluate these five categories of embodied understanding on a suite of 28 embodied spatial intelligence and planning benchmarks. The categories are foundational visual-spatial perception; 3D and multi-view spatial understanding; embodied cognition, reasoning, and planning; spatial grounding, pointing, and affordance; and visual trace and trajectory reasoning.
On this suite, PhysBrain 1.5-8B achieves a 72.5 overall score, ranking first among the evaluated open-source models and performing comparably to leading proprietary counterparts. See Evaluation Notes for details on the evaluation framework and protocol.
The table below breaks down these overall scores by benchmark, grouped into the five categories of embodied understanding.
The examples below illustrate specific tasks within each category:
- Foundational Visual-Spatial Perception: Object counting, spatial relationship reasoning, relative depth comparison, and metric size estimation.
- 3D and Multi-View Spatial Understanding: Absolute metric distance estimation, room area calculation, and egocentric motion reasoning under camera viewpoint shifts.
- Embodied Cognition, Reasoning, and Planning: Temporal past-action description, counterfactual future prediction, step-by-step goal decomposition, and task execution outcome estimation.
- Spatial Grounding, Pointing, and Affordance: Point-level object grounding, target region localization, and functional affordance identification.
- Visual Trace and Trajectory Reasoning: Waypoint generation for obstacle-free movement, object rearrangement, and contact-rich manipulation traces.
Action Prediction
Given a task instruction, current visual observations, and optional recent action history, PhysBrain 1.5 predicts the next end-effector action chunk as a compact sequence of ActionPiece tokens. A unified action codebook and vocabulary are shared across diverse control configurations and arm setups, enabling a single generalist model checkpoint to drive varied robotic platforms and manipulation settings.
Future-state Prediction
Given the current visual observation and a task instruction, PhysBrain 1.5 predicts a possible future state of the world, represented by an RGB image, a depth map, and a robot mask. The examples below illustrate predictions one second into the future across diverse robot embodiments.
Evaluation Notes
Because results reported by different works may use inconsistent metrics, we independently re-evaluate all comparison models with one canonical metric per benchmark, making the scores directly comparable. When a model provides an official benchmark-specific input format or prompt, we follow it; otherwise, we use a default task prompt and input template. For consistency across models, we standardize source-image resolution and the number of sampled video frames while retaining each model’s native preprocessing pipeline. We also follow each model’s recommended thinking configuration and point-output format whenever available.
Our evaluation is implemented with the EmbodiedEvalKit framework. First, for ERQA, EgoPlan-Bench2, SAT, ShareRobot-Trajectory, VABench-Visual-Trace, BLINK, CV-Bench, VSI-Bench, EmbSpatial, COSMOS, RoboVQA, and VLABench, we follow the toolkit’s input protocols, prompts, and evaluation metrics; for the two trajectory benchmarks, we report $100-\mathrm{NRMSE}$ so that higher values consistently indicate better performance. Second, for ten point-localization and grounding benchmarks—Part-Affordance-2K, PIOBench, PixMo-Points, PointBench, RefSpatial-Bench, RoboAfford, RoboRefit, RoboSpatial-Home, VABench-Point, and Where2Place—we retain EmbodiedEvalKit’s input protocols and prompting templates but adopt a unified micro-averaged F1 score to jointly measure prediction validity and target coverage. Third, for benchmarks outside the toolkit—ERQA-PLUS, ViewSpatial-Bench, MindCube, MMSI-Bench, 3DSRBench, and Q-Spatial-Bench—we retain their official data, task formulations, and evaluation rules. Further evaluation details will be provided in the technical report.
Quick Start
PhysBrain 1.5 uses pretrained Qwen3-VL as its backbone and extends the vocabulary with dedicated action and visual-state tokens. Its standard model interface allows you to reuse existing infrastructure for inference and post-training, including Transformers, vLLM, SGLang, LLaMA-Factory, ms-swift, and veRL.
Citation
@misc{physbrain1.5,
title={PhysBrain 1.5: From Vision-Language Models to Physical Foundation Models},
author={DeepCybo Team and Yu Bin and Haipeng Cao and Zheng Chang and Kai Chen and Youning Chen and Kailin Deng and Yichao Du and Xiaotong Fu and Haoyang Ge and Yunlong Guo and Chenliu Hao and Jiyan He and Xuguo He and Yakun Hou and Kai Hu and Cong Huang and Tuopusen Huang and Yu Huang and Hong Li and Peize Li and Shijie Lian and Xiaopeng Lin and Yun Lin and Haibao Liu and Haochen Liu and Qiuzhi Liu and Shengcai Liu and Zhiqiang Liu and Tao Luo and Peng Ren and Shuo Ren and Chaoyi Ruan and Zhaolong Shen and Yukun Shi and Qiyuan Su and Yuxuan Tian and Yining Wang and Changti Wu and Hao Wu and Xueyin Xu and Ruoqi Yang and Zhaoyang Yang and Hang Yuan and Zhaoyang Zeng and Hanwen Zhang and Ruimeng Zhang and Yao Zhang and Yibo Zhang and Yuxiang Zhang and Zhirui Zhang and Ziyi Zhang and Zubin Zheng and Zishen Zhuang},
year={2026},
eprint={2609.14973},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2609.14973},
}