Gasgoo Munich- A major early-stage deal has emerged in the embodied intelligence sector. LatentVerse, a developer of foundation models for robotics, has secured hundreds of millions of yuan in a seed round. The funding, announced on August 11, underscores the intensifying investment in foundational AI models for robotics — particularly for a company founded just this May.
Image Source: LatentVerse
LatentVerse was spun out of the embodied intelligence research lab at Tsinghua University's Institute for Interdisciplinary Information Sciences (IIIS). Founder and CEO Hu Yucheng, a PhD candidate at IIIS, studied under Chen Jianyu — the founder of ROBOTERA— and previously worked on foundation models at ByteDance's Seed unit. The team also includes talent from Peking University, Nanyang Technological University, and teams at Alibaba's Qwen and Xiaomi.
Unlike startups focusing on hardware, LatentVerse prioritizes the model. The team's prior work spans projects like DP3, HiRT, VPP, BagelVLA, and UAM, tracing a technical evolution from Vision-Action (VA) and Vision-Language-Action (VLA) to World-Action Models (WAM) and unified architectures. Notably, research on DP3 pioneered the integration of compact 3D visual representations into diffusion policies, demonstrating positive results in generalizing robot manipulation tasks.
This approach is becoming a new frontier in robot foundation models. Interest in World-Action Models has increased this year. NVIDIA's DreamZero, for instance, jointly predicts future video frames and robot actions, enabling the model to learn environmental shifts and control strategies simultaneously. In real-world tests, it has demonstrated significant capability to handle unseen tasks. NVIDIA has recently integrated DreamZero, VPP, and LingBot-VA into its WAM technical roadmap.
LatentVerse has taken this further with UTAM (Unified Tactile Action Model), an "embodied full-modality" world model that aims to integrate vision-language, world modeling, action, and tactile sensing into a single foundation. In this architecture, the vision-language module interprets tasks, the world model predicts outcomes, the action module generates control signals, and the tactile module makes high-frequency adjustments based on contact force, slippage, and object deformation.
Prioritizing tactile sensing differentiates LatentVerse from many VLA-focused competitors. While robots can now reliably handle individual tasks like picking and placing objects, the main challenge lies in maintaining stability with unfamiliar objects, continuous tasks, and complex physical interactions. Merely identifying a cup's location via camera is significantly different from detecting whether that cup is slipping from a gripper's grasp.
However, the path from architectural innovation to scaled commercial deployment is fraught with obstacles: data scarcity, computing demands, hardware compatibility, and reliability in real-world scenarios. Even NVIDIA acknowledges that while WAM promises better generalization for new tasks, understanding spatial physics and transferring skills across different robot bodies remain open problems. Figure's Helix, which uses a "slow-thinking, fast-acting" dual-system architecture, suggests the industry is still testing different approaches to balancing high-level reasoning with real-time control.
LatentVerse has already built the multimodal data infrastructure and training pipelines for UTAM, and has begun joint training of its expert modules. The new capital will be allocated to model R&D, multimodal embodied data systems, training infrastructure, and real-world robot validation. For a company months old, raising hundreds of millions is the initial cost. The main test lies ahead: whether UTAM can move beyond papers and demos to prove its ability to handle cross-scenario generalization, long-horizon tasks, and dexterous manipulation.
About "Seeds":
Gasgoo's "Seeds" column serves as a bridge connecting startups, ecosystem partners, investors, and local governments, aiming to empower the broader automotive supply chain. Since its inception, the column has focused on uncovering the companies, technologies, and business models shaping the intelligent transformation — driving the growth of industrial innovation. According to Gasgoo data, nearly every startup featured in "Seeds" has subsequently secured vital connections within the industry ecosystem.
