Blog

这里记录我在知乎专栏发布的最新技术文章，主要聚焦具身智能、机器人与多模态大模型。

知乎文章

• Reinforced Embodied Planning with Verifiable Reward for Real-World Robotic Manipulation / 《具身规划》
  Explains how the REVER framework from Xiaomi Robotics Lab and Zhejiang University fine-tunes VLM planners with verifiable rewards to keep long-horizon manipulation reliable in the real world.
  介绍 Xiaomi Robotics Lab 与浙江大学合作的具身规划研究，重点阐述如何让 VLM 从“看得懂”走向“做得对”，以及具身规划的现实瓶颈与解决思路。

• Embodied Foundation Model / 《具身基座模型》
  Provides a 2025 snapshot of representative embodied foundation models—including Cosmos-Reason1, Magma, VeBrain, RoboBrain, and Embodied-R1—and reflects on Gemini Robotics 1.5.
  概览 2025 年具有代表性的具身基座模型，包括 Cosmos-Reason1、Magma、VeBrain、RoboBrain 与 Embodied-R1，并讨论 Gemini Robotics 1.5 的最新进展。

• AgiBot GO-1: The Evolution of Generalist Embodied Foundation Model from VLA to ViLLA / 《智元机器人 AgiBot GO-1：通用具身基座模型的演进》
  Dissects how AgiBot GO-1 upgrades from VLA to ViLLA, covering capability transfer, hardware-software co-design, and lessons for generalist embodied agents.
  分析 AgiBot GO-1 如何从 VLA 升级到 ViLLA，涵盖能力迁移、软硬件协同设计，以及对通用具身智能体的启示。

• DeepSeek-V3: The Strongest Open-Source Foundation Model / 《最强开源大模型 DeepSeek-V3》
  Summarizes DeepSeek-V3’s architecture, training stack, and benchmark wins, highlighting why it sets a new bar for open-source LLM deployments.
  概述 DeepSeek-V3 的体系结构、训练栈与基准成绩，解析其为何成为开源大模型的新标杆。

• Real-Time Intelligent Systems / 《实时智能系统》
  Outlines how real-time constraints intersect with AI pipelines, from scheduling and perception to closed-loop decision-making for safety-critical robots.
  讨论实时约束如何与智能系统流程耦合，覆盖调度、感知到闭环决策，对安全关键机器人具有指导意义。