Dreamok founder Du Yu ran autonomous driving at a leading Chinese EV maker and is betting his old team's sense decide act data loop — sensing, deciding, acting, then feeding back — belongs in the bedroom.
A former head of autonomous-driving at a leading Chinese EV maker has spent the past year and a half trying to make the bed the next place his old team's data loop runs. His startup, Dreamok (智梦可), launched an AI sleep mat on JD.com in early July: a roughly 1.7-centimeter pad that lays on top of an existing mattress, watches the sleeper through the night, and adjusts its own temperature continuously across each sleep stage.
Sleep trackers sense and advise; smart mattresses tilt or warm; neither closes the loop by acting on what it sensed. Dreamok's founder Du Yu (杜宇) told 量子位 (QbitAI) that the team took its framing directly from the autonomous-driving world: perceive the environment, decide what to do, act, then feed the result back into the next decision. In a car, that loop runs at hundreds of hertz against a road. In a bedroom, the company is running the same shape of loop against a single body on a single pad, with mattress temperature as the continuously adjusted variable.
What the system actually senses is broad enough to make the comparison credible. The TopLay sensing-and-cooling mat picks up heart rate through the sleeper's back via ballistocardiography (the same signal path used in some wearable heart studies), reads breathing spectrum and body motion, and circulates temperature-controlled water through its layers. A smart hub runs on-device decisioning. A bedside companion robot called Xiaoke handles environment sensing and voice interaction, and serves as the entry point for what the company calls a "health butler" agent that follows the user past the bedroom into daytime energy management. A companion app extends that surface area further. (Sina Finance interview profile, NetEase follow-up)
The closed-loop idea is the part Du Yu wants readers to take away. Most existing sleep products, he argues in the QbitAI interview, fall into one of two buckets: traditional adjusters, which move softness, angle, or firmness, and recorders-and-suggesters, which hand the user a recommendation but leave the user to act. Treating temperature as a variable the system owns, rather than a setting the user dials once before bed, is what Dreamok is selling. The data point Du Yu returns to is the everyday failure case: a user who kicks the covers off at 3 a.m. because they got too hot, then wakes up cold an hour later when their body cools.
The public AI-plus-sleep category has visibly crowded over 2025 and 2026. Competitors 今日宜休 (Jinri Yixiu) and 格物科技 (Gewu Tech) have entered public view alongside Dreamok, and a Geekpark feature reports the US sleep-tech segment pulled in roughly $500 million across the six months before its piece ran (Geekpark context). That capital flow and that entrant count put the category on a contested footing rather than a one-off launch, and they explain why Dreamok is foregrounding paradigm transfer over feature selling.
What is not yet on the record is whether the closed-loop temperature control actually moves the outcomes the pitch depends on: sleep latency, deep-sleep ratio, night-waking frequency. The design numbers the team cites on mat thickness, a 5-micron PVDF sensing layer, and 1000-hertz sampling come from the company's own 量子位 interview, not from an independent sleep-lab study; the founder's prior employer, described only as "a leading NEV maker" (new-energy vehicle maker), is unnamed in the source. What to watch next: whether Dreamok or any of its Chinese rivals publish clinical or in-home trial data, or whether the category keeps shipping on mechanism and timing alone.