AINEWS
China's Xingchi Dynamics says the next home robot's edge will come not from gripper benchmarks but from modeling the humans it lives with.
The husband is sure the milk is in the microwave. His wife has already taken it out, put it back in the fridge, and walked away. A home robot that only watches the room sees an empty microwave. A home robot that watches the husband sees a man about to act on a wrong belief.
That second kind of robot is the bet Xingchi Dynamics is making. The Chinese embodied-AI startup has spent several years on a home robot architecture it calls PULSE. In a long interview with 雷锋网 (Leiphone) / AI 科技评论, CTO 李达 (Li Da) argued the field has been optimizing the wrong variable. "We've been training robots to grasp objects, navigate rooms, fold clothes. That's physics," Li said. "The hard part at home isn't the physics. It's the person."
PULSE's central architectural move is to split the robot's "brain" into two parallel tracks that cross-validate each other. The first track does what most embodied-AI systems already do: perceive and predict the physical environment. The second track does what almost none of them do: perceive and predict the human in the room, modeled as a vector of intent, belief, and preference. The two tracks fuse at the feature level through cross-attention, not as a bolted-on module after the fact. User state becomes a set of latent vectors, token-like, that participate in the same attention layers as the environment. "Understanding humans is architecturally equal to understanding physics," Li said. "It's not a downstream add-on. It's the other half of the world model."
The three load-bearing keywords for that user model are 意图 (intent), 信念 (belief, including false beliefs like the milk-in-the-microwave case), and 偏好 (preference, accumulated across long-term interaction in a single household). The implicit claim is that human behavior inside one home is more regular than physical-world generalization, the longer road every embodied-AI lab is on.
Li used an illustration that is not a benchmark. A grasp model trained on an 80-centimeter table, he said, fails on a 75-centimeter table. "Physical intelligence at home is the same story. It is going to be slow." The bet is that modeling one household's patterns, day after day, is a tractable near-term problem.
Deployment matches the bet. Cloud handles global iteration across the fleet. The edge handles test-time adaptation: the home robot fine-tunes locally on new-distribution data without waiting for a cloud roundtrip. Each household's robot is expected to evolve on its own track.
PULSE itself is a portmanteau: 永续 (sustainable), stitched to value alignment and continuous evolution. The English reading is a heartbeat, Li said: a robot that has a "life" and grows with the user. The product framing is 察言观色的伙伴, an attentive partner suited to a low-tolerance home. "The question at home isn't only can the robot act. It is should it act, when, and how."
The honesty floor: every architecture, design, and philosophy claim in the PULSE pitch is a first-party CTO statement. There is no independent benchmark, no peer-reviewed paper, and no third-party reproduction visible in the public record. PULSE may not be the first architecture to push user-state prediction into the world model. The dual-track framing could collapse to a single model in practice. Li's argument that human behavior in one household is more regular than physical generalization is plausible, not validated. The bet is real. The receipts are not yet.
For a cold reader: embodied AI is the field that puts intelligence inside physical robots. Theory of mind is the psychology phrase for modeling what another being believes. Xingchi Dynamics is making the case that the next home robot will not be smarter than the person it lives with; it will be better at reading that person. The 80cm-versus-75cm table is the concession inside the pitch: physics-first home robotics is the slower road, and the company is deliberately pulling back from it.
The watch item is whether edge-side test-time adaptation can compound across a real fleet. Cloud-side fleet learning is a solved engineering pattern. Per-household, on-device, lifelong user-model training is not. If PULSE can show one household's robot improving on a measurable user-prediction task over six months without cloud roundtrip, the bet moves from philosophy to product. If it cannot, the dual-track reframe reads as marketing around an unsolved problem.
The five-stage loop Li says anchors the system is a clean summary of the bet: 看清环境 (see the environment), 理解用户 (understand the user), 双轨推演 (dual-track prediction), 接受反馈 (accept feedback), 持续进化 (keep evolving). The "understand the user" stage is the one the rest of the field has been treating as decoration.