36Kr Exclusive: CUHK PhD and Former DJI Engineer Found Consumer Quadruped Robot Startup, Raises tens of millions in Angel Round Led by Zhengxuan 36氪首发 | 港中大博士、前大疆工程师创业消费级四足机器人，天使轮获正轩领投数千万元

TL;DR

• Lumisition Robotics secured nearly 50 million RMB in angel funding led by Zhengxuan Investment to develop consumer-grade quadruped robots focused on physical labor augmentation rather than companionship.
• The company targets outdoor scenarios for detached home owners and seniors, utilizing a four-layer intelligent architecture: perception, motion, intent, and role intelligence.
• Key algorithmic innovations include the LMV framework for learning movement from raw video, semantic topology maps for navigation, and interactive navigation that allows the robot to manipulate objects to create paths.
• Hardware cost reduction is achieved through multi-disciplinary optimization of joint modules (reducing parts from 18 to 11) and model distillation to lower computational requirements.
• The product roadmap includes outdoor testing in October 2026, a formal launch at CES 2027, and mass production scheduled for April-May 2027.

Why It Matters

This development marks a strategic shift in the consumer robotics market, moving away from emotional companionship toward practical utility and physical augmentation, which addresses specific pain points like aging populations and outdoor labor. The integration of advanced reinforcement learning and semantic navigation into a cost-effective hardware platform demonstrates a viable path for scaling quadruped robots beyond industrial use cases into mainstream consumer adoption.

Technical Details

• Four-Layer Intelligent Architecture: Comprises Perception Intelligence (environmental info), Motion Intelligence (stable movement via RL), Intent Intelligence (Agent-based decision making for speed vs. safety), and Role Intelligence (dynamic functional switching and modular adaptation).
• Advanced Motion Control: Utilizes a Terrain Map combined with online state estimation and structured foot rewards for robust obstacle crossing. The LMV (Learn from Motion Video) framework enables policy learning directly from raw video data, reducing reliance on mocap or simulation.
• Semantic Navigation: Replaces traditional geometric mapping with semantic topology graphs for "decision-based routing." An interactive navigation framework allows the robot to actively manipulate environmental objects (e.g., pushing doors or moving chairs) to navigate dynamic, semi-structured spaces.
• Hardware Optimization: Joint module design was optimized from 18 to 11 parts to reduce manufacturing costs. Navigation systems use low-cost camera and LiDAR combinations, supported by model distillation to minimize high-performance computing needs.

Industry Insight

• Differentiation Strategy: Companies should focus on solving tangible "last-mile" physical problems (like carrying loads or assisting mobility) rather than competing solely on emotional value, targeting niche demographics such as the "garage culture" enthusiasts and the elderly.
• Data Flywheel Potential: Quadruped robots serve as natural data collection terminals for complex non-structured environments; building a proprietary dataset of terrain interactions and friction data can create significant long-term competitive moats.
• Cost-Performance Balance: Successful consumer robotics requires systemic cost optimization across the entire supply chain, including algorithm-driven hardware simplification and sensor selection, to achieve price points accessible to average consumers while maintaining robust performance.