In the ongoing debate about embodied intelligence for home robots, one question persists: what makes people keep a robot at home? In factories, the logic is simple—precision, stability, and productivity. But in homes, the tolerance for error is near zero. A single mistake or misinterpretation can lead to returns or abandonment.
This reality has split the industry into two paths. One camp continues to push physical intelligence, striving for robots that understand the physical world perfectly. The other steps back to first ask: how can a robot understand people well enough to live with them? Xingchi Power (星炽动力) firmly belongs to the latter.
In a recent interview with AI Tech Review, CTO Li Da elaborated on the company's PULSE architecture. Li, a PhD from the Institute of Automation, Chinese Academy of Sciences (under academician Tan Tieniu), spent over a decade in visual navigation and continual learning before bringing his expertise to home robotics. His core thesis: "Today's embodied brains focus on 'environment dynamics'—but we must also infer the user's state."
PULSE splits the embodied brain into two tracks: one for physical environment perception and reasoning, the other for user state perception and reasoning. These tracks don't operate in isolation; they cross-validate at the feature level. Physical information tells the robot "what's happening in the environment," while user state tells it "what's happening with the person." Combined, they produce more reliable decisions.
In a home, people are not just objects in the environment—they are the center of interaction. If a robot's cognitive model only includes tables, cups, and sofas, but ignores the person sitting on the sofa, it remains a passive command-line tool. But homes include elderly, children, and others who may not articulate clear commands. How does PULSE make robots understand people? The answer lies in three keywords: intention, belief, and preference.
Intention is about reading what the user says or does. Belief is more subtle: Li gives an example where a husband thinks milk is still in the microwave, but the wife has already taken it out. The husband holds a false belief. A robot that only tracks physical objects sees an empty microwave and cannot predict the husband's next action based on a mistaken premise. But if the robot can represent the user's belief state, it can proactively remind: "The milk has been taken out." This correction capability is true home intelligence.
Preference is longer-term, covering habits, boundaries, and privacy, accumulated through continuous interaction. These three dimensions form a user model that is not an add-on module but embedded directly into the world model's conditional input. User state is represented as latent feature vectors, participating in cross-attention computation alongside environmental features.
This technical choice reflects a clear value judgment: understanding humans and understanding the physical world should have equal status in the cognitive architecture—perhaps even with humans prioritized. Why? Because the physical world is too complex for near-term generalization. A grasping model trained on 80cm tables may fail on 75cm tables. Universal physical intelligence across all homes, objects, and lighting conditions is not imminent.
But human understanding models are different. While humans are complex, individual behavior within a single home is patterned—when they drink water, which cup they use, their tone of voice. If a robot can first understand "this person" in a specific home, its usability rises quickly. Higher usability drives more frequent use, generating real user feedback data that can further improve the embodied brain through tight coupling with the robot's own hardware.
PULSE's end-cloud deployment strategy reflects this: the cloud handles global iteration, while the edge runs test-time adaptation. When encountering new distribution data, the robot tunes locally without waiting for cloud round-trips. This design naturally fits home scenarios: every home is different, and every robot should develop its own understanding model.
Ultimately, physical intelligence solves "can it be done?" while human understanding solves "should it be done, when, and how?" In the low-tolerance home environment, the latter may deserve higher priority. As Li Da puts it: "The key to embodied intelligence in homes is not making robots more like universal tools, but more like perceptive companions."
Source: 雷峰网.