UCLA PhD Team Founding Humanoid Robot Base Model, Secures Nearly 500 Million Yuan Angel++ Round | Hard Krunch Exclusive UCLA博士团队创业做人形机器人基础模型，拿下近5亿元天使++轮融资｜硬氪首发

TL;DR

• Delta Intelligence raised nearly 500 million RMB in a Series++ round to iterate humanoid foundation models, accelerate data collection equipment production, and expand the core R&D team.
• The company focuses on developing native general-purpose Humanoid Foundation Models (HFMs) for full-body coordinated manipulation, aiming to advance Physical AGI through humanoid robots.
• Delta's architecture features a "Brain + Cerebellum + Force-Position Hybrid" three-layer collaborative system with distinct roles: Brain handles perception and planning using a 3D world engine, Cerebellum manages low-level motor control via reinforcement learning, and the hybrid layer ensures smooth execution.
• They utilize a proprietary whole-body panoramic data acquisition system capturing full human skeleton motion and high-fidelity 3D scene data from first-person video streams, enabling comprehensive interaction modeling.
• A standardized adaptation process allows rapid deployment across different robot platforms by separating simulation-based cerebellum training from real-data brain fine-tuning, minimizing hardware wear and retraining costs.

Why It Matters

This development represents a significant leap toward practical humanoid robotics by addressing critical limitations in current embodied AI systems—specifically the gap between high-level task planning and low-level physical control in complex environments. By pioneering native 3D spatial understanding and decoupling cognitive planning from motor coordination through its layered architecture, Delta Intelligence offers a scalable solution that could accelerate industrial adoption of autonomous humanoid workers capable of navigating unstructured spaces while performing delicate manipulations simultaneously.

Technical Details

• Three-Layer Architecture: Implements a hierarchical control structure where the Brain module uses a self-developed 3D world engine for long-term task planning and environmental reasoning; the Cerebellum employs neural networks trained via massive simulation reinforcement learning to generate precise joint torque commands at hundreds of Hz; and the Force-Position Hybrid layer outputs compliant motions ensuring safe physical interaction.
• Native 3D Representation: Unlike conventional approaches relying on 2D visual features prone to depth ambiguity, Delta’s system processes raw point clouds and Gaussian splatting representations directly within its world engine, enabling unambiguous geometric reasoning essential for tasks like door opening or stair climbing.
• Whole-Body Data Collection System: Proprietary pure-vision capture rig records synchronized kinematic trajectories of all major joints during human demonstrations alongside incremental reconstruction of surrounding geometry from egocentric camera feeds, producing dual datasets comprising full-body skeletal sequences and metric-scale obstacle maps.
• Hybrid Training Strategy: Separates learning objectives between modules—the Brain exclusively trains on authentic teleoperated interactions involving multi-point contact forces impossible to simulate accurately, whereas the Cerebellum leverages physics engines for infinite trial-and-error balance optimization before minimal real-world calibration.
• Cross-Humanoid Adaptation Pipeline: Four-stage workflow including motion retargeting under dynamic constraints preserving force distribution patterns during transfers between diverse chassis configurations, followed by independent refinement paths optimized per platform characteristics without requiring complete retraining cycles.

Industry Insight

The emergence of specialized foundation model providers like Delta Intelligence signals an impending shift away from vertically integrated monolithic designs toward modular ecosystems where software layers become agnostic underlying hardware specifics—a trend likely driven by increasing fragmentation among robot manufacturers seeking flexibility rather than proprietary lock-in effects. As competition intensifies around achieving reliable dexterity outside controlled labs, companies investing heavily in robust sensory fusion techniques coupled with efficient sim-to-real transfer mechanisms will hold decisive advantages when scaling deployments beyond niche pilot programs into mainstream manufacturing workflows. Furthermore, establishing open standards for interoperability between these foundational algorithms and various actuator architectures may soon define new battlegrounds for influence over how quickly humanity transitions from observing robots merely walking around to actually employing them as productive members of our daily lives both professionally domestically alike.

TL;DR

• 德塔智能完成近5亿元天使++轮融资，专注于人形机器人基础模型（HFMs）研发，旨在实现通用的人形全身协同操作。
• 团队采用“大脑+小脑+力位混合”三层原生协同架构，结合三维世界引擎和全身全景数据采集技术，提升机器人在复杂环境中的理解与操作能力。
• 公司创始人均为UCLA博士，具备丰富的学术与产业背景，已在人形机器人领域取得多项重要研究成果。
• 通过与多家头部人形机器人厂商合作，推动技术在真实工业场景中的工程化落地与验证，加速具身智能的产业化进程。

为什么值得看

这篇文章详细介绍了德塔智能在人形机器人基础模型领域的技术创新和商业模式，对于关注具身智能和人形机器人发展的从业者来说，具有重要的参考价值。通过了解其技术路径和行业布局，可以更好地把握未来机器人技术的发展趋势和市场机会。

技术解析

• 三层原生协同架构：德塔智能采用了“大脑+小脑+力位混合”的三层架构，其中大脑负责环境感知和任务规划，小脑负责底层控制与平衡，力位混合层则输出柔顺的控制信号，确保机器人在复杂环境中的稳定运行。
• 三维世界引擎：为了克服二维视觉表征的局限性，德塔智能构建了原生的三维世界引擎，能够直接处理点云和高斯泼溅等三维场景表示，实现对环境的精准理解和推理。
• 全身全景数采设备：自研的数采设备可以同步捕捉人体全维度关节运动轨迹，并基于第一视角视频流实时重建全局三维场景，为模型训练提供高质量的数据支持。
• 分层数据采集策略：针对不同开发阶段，德塔智能设计了分层数据采集方案，包括搭载机器人本体的遥操作采集和无本体动捕模式，以平衡数据规模、成本与精度。
• 跨本体适配流程：针对不同类型的人形机器人本体，德塔智能建立了一套标准化的适配流程，涵盖数据采集、运动重定向、仿真强化学习和微调等环节，实现了模型在不同硬件平台上的快速迁移。

行业启示

• 技术融合与创新：德塔智能的成功表明，将前沿的人工智能技术与机器人硬件深度融合，是突破现有瓶颈、实现高效人机协作的关键方向。
• 生态合作的重要性：通过与多家头部人形机器人厂商的合作，德塔智能不仅加快了技术的迭代速度，还促进了整个产业链的发展，形成了良好的生态系统。
• 数据驱动的未来：随着数据采集技术和算法模型的不断进步，数据将成为推动人形机器人发展的核心资源，企业应重视数据积累和管理，以提升竞争力。