Can Robots Really Understand Humans? Inside the VLIA Intent‑Driven Embodied AI Model

In 2026, embodied intelligence reaches a turning point as humanoid robots shift from industrial ROI validation to household and consumer scenarios. While most companies focus on making robots "do tasks," DeepStar (深穹星核) argues that the real challenge is whether robots can truly understand the people they serve.

Non‑consensus stance: The core proposition for consumer‑oriented embodied AI is not task execution but intention comprehension. Human language is noisy and incomplete; a simple "OK" may hide preferences, and silence can be a crucial signal. Vision, language, and cloud models only address surface interaction; only by reconstructing "why a person says something, what they truly need, and how they wish to be responded to" can a robot claim to understand humans.

VLIA model: DeepStar introduces VLIA (Vision‑Language‑Intention‑Action), the first end‑side embodied interaction model that embeds intention understanding at its architectural core. Vision and language are no longer mere input channels, nor is action just an output command; the model continuously restores real intent and predicts interaction trajectories. Human faces serve as the highest‑bandwidth entry point, embedding motive, emotion, relationship, and feedback into the model.

Data as the ceiling: Public video and speech data lack structured intent annotations. DeepStar partners with leading short‑drama studios and female‑oriented game developers to acquire seed data that already contains complete intent structures. This seed data is processed through a proprietary pipeline, generalized to massive public interaction logs, and transformed into intent‑supervised trajectories.

Cross‑disciplinary expertise: Founder Yu Zhenbo, a three‑time ImageNet competition champion with large‑scale training experience, brings end‑side development expertise from Huawei HiSilicon. Core team members from Tencent and miHoYo contribute deep knowledge of player intent modeling, forming the technical foundation for VLIA’s intention understanding.

Hardware integration: The Nova S1 humanoid robot, equipped with the VLIA brain, targets home interaction and collaboration. Facial design is treated as a high‑bandwidth emotional interface; DeepStar collaborates with Shanghai Ninth Hospital’s plastic surgery department to build a medical‑grade, parametric, 3D‑scanned facial pipeline that meets anatomical, engineering, and aesthetic standards.

Strategic partnerships: Collaboration with the StarFlash Alliance (星闪联盟) provides low‑latency wireless connectivity and hardware co‑design, while academic advisor Academician Zhang Wenjun guides the definition of embodied intelligence standards.

Scaling intent data: The scaling roadmap consists of five steps: (1) Intent Seed Data from drama and games, (2) Annotation Model for intent labeling, (3) Large‑scale processing of video, dialogue, and interaction logs, (4) Generation of intention‑supervised trajectories, and (5) Training VLIA on these trajectories. The focus is on expanding "learnable human intent trajectories" rather than merely increasing raw video volume.

Personalization mechanism: To address long‑term companionship challenges such as behavior drift and personality inconsistency, DeepStar introduces OrthLoRA, a technique that isolates stable core capabilities from mutable personal preferences within the model’s parameter space, enabling continuous, consistent personalization.

Technical moat: DeepStar’s three‑layer moat—medical‑grade facial aesthetics, exclusive structured intent data, and scalable data‑infra engineering—creates a closed loop that is difficult for competitors to replicate, securing a competitive advantage in the emerging household robot market.