PhysBrain 1.5: Chinese Physical AI Breakthrough
The Physical AI Race Heats Up
Physical AI is becoming the most critical track in global AI for 2026. While large models have spent three years pushing "can talk, can think, can work" to the extreme, the industry is now asking: Can models enter the real world? Can they take action?
Recent experiments have put this question front and center:
- Gemini Robotics 2 (Google DeepMind, July 2026): Based on Gemini 3.5 Flash, uses "embodied reasoning" brain ER 2 to make robots "watch videos and plan next steps"
- GPT-6 Astra (OpenAI): Pushed the ceiling of general intelligence higher with breakthroughs in reasoning, coding, and multimodal understanding
- Robocurve Experiment: Deployed GPT-6 Astra on a real robotic arm — completed 19/20 "put blocks in bowl" tasks, while Claude Fable 5.1 only completed 8
But top LLMs still have a gap in physical manipulation. On millimeter-level precision tasks like "blue puzzle piece alignment insertion," GPT-6 Astra's success rate drops from 95% to 10%.
Enter PhysBrain 1.5
On September 9, 2026, Chinese company DeepCybo (深度机智) released PhysBrain 1.5, a physical foundation model that achieves 72.5 average score across 28 public benchmarks — ranking #1 among open-source models, just 1 point behind top closed-source models GPT-6 Astra (73.3) and Gemini 3.6 Flash (73.0).
Key Achievements
| Metric | Result |
|---|---|
| Open Source Rank | #1 |
| Average Score (28 benchmarks) | 72.5 |
| Open Source #1 | 14 benchmarks |
| Open Source #2 | 10 benchmarks |
| vs. Hy-Embodied-VLM-1.0 (66.0) | +6.5 points |
| vs. RynnBrain 1.1 (63.1) | +9.4 points |
Two Model Sizes
- PhysBrain 1.5-2B: 66.6 average — exceeds all non-PhysBrain open models
- PhysBrain 1.5-8B: 72.5 average — tops open source, nears closed-source leaders
Both models are fully open-source with Apache 2.0 license.
The Physical Loop Architecture
PhysBrain 1.5 is built around a unified architecture called Physical Loop:
- Observe the world — perceive the environment
- Understand space and tasks — comprehend spatial relationships
- Judge action consequences — predict outcomes
- Execute — take physical action
- Correct based on feedback — self-correct from new observations
This is a continuous, stable closed-loop system that self-corrects based on feedback.
Three Core Capabilities
- Embodied Understanding: Knows "what is happening now" — outputs structured spatial coordinates and trajectory targets
- Action Generation: Knows "what to do next" — learns from human video, adapts to different robot types
- Future State Prediction: Foresees "what happens after" — predicts 1-second physical state in RGB, depth, and robot mask
The Human Learning Approach
DeepCybo's key insight: Physical intelligence cannot be achieved by translating internet text into actions. Humans grow operational skills through repeated "see — try — get feedback — correct."
The team built Ego360, a human panoramic real-data system, using full-body pose, hand movement, and task-level voice from panoramic video. Physical pre-training supervision comes entirely from human interaction videos.
Open Source & Community
- Project Page: deepcybo-physai.github.io/PhysBrain-1.5
- Hugging Face: huggingface.co/collections/DeepCybo/physbrain-15
- Eval Kit: github.com/DeepCybo-PhysAI/PhysBrainEvalKit
- Demo: huggingface.co/spaces/hugging-apps/physbrain1-5-8b-demo
Within 3 days of launch, PhysBrain 1.5 received 3,000+ downloads on Hugging Face.
Conclusion
PhysBrain 1.5 demonstrates that physical foundation models are not a short-term trend but a long-term strategic investment. By focusing on human learning patterns and building a complete data-model-body loop, DeepCybo has positioned PhysBrain at the forefront of global physical AI.
Tags: AI, Robotics, PhysicalAI, OpenSource, DeepLearning
Top comments (0)