AI News AI资讯 22h ago Updated 18h ago 更新于 18小时前 50

Mo Shen Intelligence Completes Nearly 100 Million RMB Pre-A Round追加 Financing 硬氪首发 | 复旦教授、前英特尔首席科学家做端侧具身大脑,「眸深智能」完成近亿元Pre-A轮追加融资

Motion Brain (眸深智能) secured nearly 100 million RMB in additional Pre-A funding, with valuations surging tenfold in the first half of 2026, driven by rapid commercialization and technical breakthroughs. The company pioneered a "World Motion Model" approach, converting continuous robot actions into discrete tokens (MotionGPT), enabling zero-shot generalization and reducing reliance on massive real-world robot data. Their STI-WM model utilizes a novel data recipe (80% internet video, 10% motion cap 眸深智能完成近亿元Pre-A轮追加融资,估值半年增长超10倍,核心团队具备海思、英特尔、英伟达背景。 提出“世界动作模型”路线,首创将连续动作离散化为Token(MotionGPT),实现Zero-Shot泛化与长序列规划。 STI-WM模型采用80%互联网视频+20%专业数据训练范式,真机数据需求降低90%,动作准确度提升至99%。 通过模型压缩与国产芯片适配,端侧推理成本从20万降至1万,延迟降至10ms,并具备离线自进化能力。 已实现商业化营收,覆盖工业、物业、环卫场景,目标成为具身智能领域的“超级通用大脑”。

75
Hot 热度
68
Quality 质量
72
Impact 影响力

Analysis 深度分析

TL;DR

  • Motion Brain (眸深智能) secured nearly 100 million RMB in additional Pre-A funding, with valuations surging tenfold in the first half of 2026, driven by rapid commercialization and technical breakthroughs.
  • The company pioneered a "World Motion Model" approach, converting continuous robot actions into discrete tokens (MotionGPT), enabling zero-shot generalization and reducing reliance on massive real-world robot data.
  • Their STI-WM model utilizes a novel data recipe (80% internet video, 10% motion capture, 10% real machine data) to cut true-machine data requirements by 90% while achieving 99% action accuracy.
  • Through aggressive model compression and adaptation to domestic Chinese chips (e.g., Ascend, Horizon), they reduced inference latency from 200ms to 10ms and deployment costs from 200,000 RMB to 10,000 RMB.
  • The team has achieved significant commercial traction with revenues reaching tens of millions of RMB in H1 2026, serving clients in industrial inspection, property management, and sanitation, while maintaining an offline self-evolving capability for edge devices.

Why It Matters

This case highlights a critical shift in Embodied AI from purely cloud-based VLA models to efficient, edge-native solutions that prioritize physical execution and cost-effectiveness. For practitioners, it demonstrates how tokenizing continuous control signals can solve long-horizon planning issues and how heavy optimization for specific hardware ecosystems (like China's domestic chip supply chain) can create sustainable competitive moats in a resource-constrained environment.

Technical Details

  • Action Tokenization (MotionGPT): Discretizes human poses into ~3,000 "action primitives," allowing the model to predict action sequences like language tokens, facilitating zero-shot generalization for unseen tasks.
  • STI-WM Architecture: A spatiotemporal unified world motion model that integrates spatial structure, temporal evolution, physical consistency, and execution robustness, trained primarily on unlabelled internet videos filtered with physical priors.
  • Data Efficiency Strategy: Replaces heavy real-world data collection with a hybrid dataset strategy, leveraging vast amounts of weakly supervised video data for physics learning and minimal high-quality motion capture/real-machine data for calibration.
  • Edge Optimization & Hardware Adaptation: Achieved a 62x reduction in computational load through token pruning and quantization, adapting models to domestic NPUs (Ascend 310/910, Horizon, SuanYuan S60) to enable low-latency (10ms) inference on edge devices without cloud dependency.
  • Offline Self-Evolution (T²MB): Implements a Task*Task Motion Brain framework allowing robots to optimize performance locally through interaction without needing to upload data or return to the factory for updates.

Industry Insight

  • Supply Chain Sovereignty as a Tech Advantage: The successful integration of AI models with domestic Chinese silicon suggests that hardware-software co-design is becoming a primary barrier to entry, especially in regions facing compute restrictions.
  • Shift from Data Scale to Data Quality & Synthesis: The industry is moving away from the "more data is better" paradigm toward synthetic data generation and physics-informed learning from existing video repositories, drastically lowering the cost of training embodied agents.
  • Consolidation in "Brain" Providers: As hardware form factors diversify, we expect a consolidation where a few key players provide universal "embodied brains" while hardware manufacturers focus on specialized mechanical designs, mirroring the smartphone OS vs. hardware dynamic.

TL;DR

  • 眸深智能完成近亿元Pre-A轮追加融资,估值半年增长超10倍,核心团队具备海思、英特尔、英伟达背景。
  • 提出“世界动作模型”路线,首创将连续动作离散化为Token(MotionGPT),实现Zero-Shot泛化与长序列规划。
  • STI-WM模型采用80%互联网视频+20%专业数据训练范式,真机数据需求降低90%,动作准确度提升至99%。
  • 通过模型压缩与国产芯片适配,端侧推理成本从20万降至1万,延迟降至10ms,并具备离线自进化能力。
  • 已实现商业化营收,覆盖工业、物业、环卫场景,目标成为具身智能领域的“超级通用大脑”。

为什么值得看

本文揭示了具身智能从“重数据依赖”向“轻量化、低成本端侧部署”转型的关键技术路径,为行业提供了可复制的降本增效方案。其“动作Token化”及“软硬协同”策略不仅解决了机器人泛化难题,更在供应链安全背景下探索了纯国产算力适配的可行性。

技术解析

  • 动作Token化架构:将人体姿态拆解为约3000个“动作基元”,类比LLM的Token机制,使机器人能通过组合基元理解未见过的指令,实现Zero-Shot泛化,发表于NeurIPS 2023。
  • STI-WM时空一体模型:基于MLD和MotionGPT迭代而来,融合空间结构、时间演化、物理一致性和执行鲁棒性,摆脱对海量真机数据的依赖,仅需10%真机数据即可完成闭环对齐。
  • 端侧优化与国产适配:通过蒸馏、量化等技术将千亿参数压缩至百亿级,参数量减少75%,推理延迟从200ms降至10ms;全面适配昇腾、燧原等国产芯片,实现全流程自主可控。
  • T²MB离线自进化机制:支持机器人在端侧无需联网即可通过本地交互持续优化性能,任务执行准确度最高提升25%,解决现场Bug需返厂升级痛点。

行业启示

  • 数据效率革命:利用互联网视频预训练结合少量真机校准的新范式,大幅降低具身智能的数据获取门槛和成本,加速模型迭代周期。
  • 端侧算力国产化机遇:随着英伟达限制加剧,具备算子级适配能力的团队将成为关键基础设施提供商,国产芯片生态与AI算法的深度耦合是必然趋势。
  • 产业分工明确化:具身智能领域将形成“通用大脑供应商+垂直本体制造商”的分层格局,初创公司应聚焦核心算法与软硬协同,避免重复造轮子。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Robotics 机器人 Funding 融资 Deployment 部署