At the 2026 World Robot Expo (WRC), held from August 19-23 at the Beijing Yichuang International Convention and Exhibition Center, a special session on "AI Large Models Empowering Robots and Embodied Intelligence: A New Industry Paradigm" took place. Li Zhanbin, founder and CEO of Xingcan Intelligence, attended and delivered a speech detailing the company's latest advancements.
Li explained that Xingcan Intelligence sends its robots directly into homes for data collection, employing a data-engine approach similar to that used in autonomous driving. This method enables low-cost, high-quality, and multi-dimensional symbiotic data acquisition. He emphasized that the company's hardware philosophy is rooted in minimalism and pragmatism, integrating model development with hardware to avoid wasting a single dollar on unnecessary components. "We don't build humanoids for their own sake," he stated.
Introducing the company's focus, Li described the "5-kilometer living circle" as the primary life scenario. Within this space, the company combines home mobility and companionship, covering the journey from bedrooms to indoor and outdoor areas, communities, and public urban spaces. He noted that each era has its specific robot products for home life, from early low-computing products like vacuum cleaners to the humanoid robots popularized after the Spring Festival Gala. Xingcan, however, is transitioning toward family companionship using mobility and companion wheelchairs as its carrier.
Drawing a parallel between autonomous driving and embodied intelligence, Li highlighted the importance of data collection methods. He argued that existing methods, such as ego-centric, teleoperation, and simulation approaches, have significant flaws. Instead, Xingcan uses its robots to enter homes directly for collection, following the autonomous driving data engine model to achieve a low-cost, high-quality, and multi-dimensional symbiotic collection process. The company has built an integrated data-collection hardware system that is itself a consumer product, entering homes on a wheelchair platform to capture 360-degree, full-dimension, high-fidelity data from ego, environment, and other perspectives, ensuring a robust and reusable data foundation for multiple scenarios.
Li then introduced the company's XcanBrain platform, which leans toward engineering practicality rather than pure research. Unlike the world model route pursued by some, Xingcan's approach offers stronger real-world applicability. The company has developed a layered decision network that integrates end-to-end models, VLA (Vision-Language-Action), and world models. This network activates different models based on specific scenarios, achieving performance comparable to chips costing nearly 10,000 yuan on hardware priced under 2,000 yuan, thereby enabling large-scale, computationally intensive models to be deployed on edge devices for practical applications.
Reflecting on the conference's themes, Li noted the resonance around spatial intelligence, referencing presentations on topics like "Galaxy" and "Corona." He emphasized that spatial intelligence represents the world representation of long-term agents, an area Xingcan has researched deeply, drawing on his personal background in autonomous driving maps and data. The company has also explored long-term memory, which, along with world representation and task execution, forms the three continuities necessary for a truly cognitive long-term agent.
Addressing the challenge of memory in 3D, Li pointed out that while large language models like DeepSeek and Kimi have extensive historical memory in 2D text, 3D data remains scarce. LiDAR has only become prevalent in the last decade, leaving most datasets lacking spatial intelligence, particularly 3D model data. Xingcan addresses this gap through its own data collection efforts, solving the problem of missing spatial data.
The company has developed two core capabilities: World2Mind and NavMind. World2Mind extends the autonomous driving process, using multi-view and point cloud multimodal information to create a mental map. NavMind then leverages this mental map for navigation, forming a plug-and-play external cognitive mapping sandbox. This approach allows for better internalization of mental navigation into structured reasoning, improving success rates from planning to real-world execution.
Li reiterated the company's hardware philosophy of defining minimalist and pragmatic forms, combining model development with hardware to avoid waste. "We don't build humanoids for their own sake," he repeated, emphasizing the focus on two primary scenarios: a Life Intelligence OS, currently in development and expected to be introduced industry-wide in the second half of next year, and mobility scenarios covering home interiors to urban public spaces, plus interaction scenarios centered on mobility and companionship. The company is also collaborating with multiple brain-computer interface (BCI) firms, both invasive and non-invasive, to integrate intent signals from BCI with the robot's environmental perception, enhancing execution smoothness and alignment with the body's needs.
During the speech, Li unveiled a new product, the Mate X. This product incorporates spectral technology previously costing tens of thousands of yuan, now achieved at a cost of around a hundred yuan. It complements existing LiDAR or RGBD sensors that capture XYZ information and visual semantic and topological data, advancing the company's progress in the material modality and further pushing multimodal capabilities. The core six capabilities include active scene perception and navigation, dynamic obstacle avoidance, intelligent following, adaptive pre-balancing, emotional companionship, and embodied companionship.
Founded in March of last year, Xingcan Intelligence has already received several awards and media recognition. Li concluded by thanking the audience and expressing gratitude for the attention the company has received.