LatentVerse bets on a model that can feel what it grips

Most of China’s embodied-intelligence start-ups are, in the end, robot makers. LatentVerse is a model maker. Its chief executive, Hu Yucheng, is a PhD candidate at Tsinghua University’s Institute for Interdisciplinary Information Sciences, and the company announced a seed round of several hundred million yuan on 11 August 2026.

Background

Hu was born in 2001, according to 36Kr’s account of the round, which was originally published by the WeChat account Intelligent Emergence. He studied under Chen Jianyu, founder of the robot maker Robot Era, and then worked on foundation models at ByteDance’s Seed unit, where he led the BagelVLA series of robot models. 36Kr reports that he proposed a method called PAD in early 2024 and a video-action model called VPP in November of that year, which it says NVIDIA later built on for a system called DreamZero. He resigned from Seed in April 2026.

LatentVerse was spun out of the embodied-intelligence research lab at his institute and was, in Gasgoo’s words, “founded just this May”. The team includes people from Tsinghua, Nanyang Technological University, Peking University, Alibaba’s Qwen team and Xiaomi. Co-founder Zhang Jianke led earlier projects including HiRT, UP-VLA and VLM4VLA, while another co-founder leads dexterous-hand algorithms and motion control.

The product

The first model is UTAM, the Unified Tactile Action Model. It combines four expert modules: vision and language, a world model, action, and a high-frequency tactile channel. The tactile module handles fine adjustments from contact force, slippage and object deformation, Gasgoo reports. The company rejects the label of a “world model company”, describing its approach as a native embodied world model, and, as Gasgoo puts it, “unlike startups focusing on hardware, LatentVerse prioritizes the model.”

Data and cost

36Kr says the company plans a robot-free collection strategy, using a self-developed data glove and partnerships with data-collection bases. It aims to cut the cost of collection from more than 1,000 yuan an hour to under 100 yuan an hour, and to accumulate hundreds of thousands of hours of data within one to one and a half years. A first foundation model of 16 billion parameters is expected within a quarter. These are targets stated by the company.

The round

36Kr names GL Ventures as lead, with Qingliu Capital, AgiBot, Robot Era, InnoAngel and other funds. Gasgoo does not name investors and says the exact amount was not disclosed. The overlap with Robot Era, Hu’s former supervisor’s company, and AgiBot is notable: two robot makers have put money into a firm that supplies the model rather than the machine.

VnExpress, in October 2026, placed Hu among Tsinghua doctoral students running start-ups, saying the company raised its seed round in less than three months.

The caution

Gasgoo is plain about what comes next: the main test lies ahead, it writes, in whether UTAM can move beyond papers and demonstrations. A seed round of this size on a company a few months old is a bet on a research record, not on a product. Hu’s record is strong. Whether tactile models turn out to be the missing piece of the robot stack is a question the lab, and now the company, still has to answer.