On October 20, the inaugural 2026 Physical AI and World Model Conference opened at the Nanjing International Expo Center. The event featured a top-tier lineup, including Turing Award winners Yann LeCun and Andrew Yao via video link, and keynote speeches by academicians Shen Xiangyang and Yu Haibin. At this premier global stage for physical AI, Qianjue Technology, a Tsinghua-affiliated company founded three years ago, unveiled its core technology: the Polibrain predictive world model, a foundation for embodied intelligence in the real physical world. This breakthrough is based on research into "latent variable identifiability" conducted by Qianjue in collaboration with Tsinghua University, which was published approximately one year ahead of similar work by LeCun's team.
The Robot Industry Has No "ChatGPT Moment"
While capital has flooded the embodied intelligence sector, high-profile humanoid robots often fail when faced with real-world variables like changing lighting or unexpected obstacles. Gao Haichuan, founder of Qianjue, noted that there is no "ChatGPT moment" for robotics because the physical world is irreversible—unlike digital text or code, physical errors carry high costs. The scarcity of large-scale physical interaction data remains the primary bottleneck.
Two Divergent Paths: Generative vs. Predictive
As "world models" become the next major battlefield, a divide has emerged between two approaches:
- Generative Route: Similar to Sora, this approach attempts to predict every pixel of future frames. Gao argues this is inefficient and prone to "physical hallucinations" (e.g., objects disappearing or passing through walls), as it wastes computing power on irrelevant background details.
- Predictive Route: This is the path chosen by Polibrain. Instead of rendering pixels, it focuses on latent representation spaces, calculating causal variables like object movement and interaction relationships. This approach prioritizes decision-making certainty over visual fidelity, using minimal computing power to achieve high efficiency.
A Mathematical Foundation Ahead of the Curve
In 2025, the Qianjue-Tsinghua team published "When Do Neural Networks Learn World Models?" at ICML, addressing how neural networks recover the true latent variables of the physical world—a year before LeCun's team explored similar territory. Their latest work introduces "Polynomial Representations" and "Effective Degree (ED)," a metric that acts like a "CT scan" for neural networks. By quantifying the smoothness of the model's internal logic, they can penalize complexity during training, forcing the model to focus on essential physical laws rather than memorizing noise.
Causal Memory and Decoupled Architecture
Polibrain utilizes a "counterfactual inference" causal memory mechanism, allowing the robot to retain only critical historical information rather than storing every second of video. This ensures consistency in the face of occlusions or sensor delays. Furthermore, Qianjue has decoupled the "brain" from the "body," allowing the same Polibrain core to be adapted to nearly 50 different robot forms, from industrial arms to quadruped platforms. By focusing on being a pure "brain infrastructure" provider, Qianjue avoids ecosystem conflicts with hardware manufacturers, a strategy that has already enabled the accumulation of massive real-world scenario data.
