david
Joined March 2026
Articles (27)
DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence
April 28, 2026
DeepSeek-V4系列预览版论文中文翻译：1.6T参数MoE模型，支持百万Token上下文，引入CSA/HCA混合注意力、mHC超连接和Muon优化器。
计算机视觉的惨痛教训（The Bitter Lesson of Computer Vision）
April 27, 2026
Vincent Sitzmann 关于计算机视觉未来的深度思考：传统的中间表示（如三维重建、分割掩码）将变得过时，计算机视觉的未来是作为端到端感知-动作循环的一部分。
《VGGT》与《SwiftVGGT》深度解读：视觉几何基础模型的统一多任务范式
April 27, 2026
深度解读 CVPR 2025 Best Paper VGGT 及其后续工作 SwiftVGGT。VGGT 通过单次前向传播同时输出相机参数、深度图、点云和点跟踪，精度超越传统优化方法。SwiftVGGT 在此基础上通过单步 SVD 和内置回环检测，将大规模场景重建速度提升 3 倍，且无需任何训练。
具身智能公司：本体-小脑路线 vs 大脑优先路线
April 24, 2026
A deep dive into the two major strategic paths in the embodied AI industry: Cerebellum-First vs. Brain-First, and the companies leading each camp.
《LingBot-Map: 用于实时三维重建的几何上下文 Transformer》深度解读
April 24, 2026
LingBot-Map 是一种基于 Transformer 的前馈式三维基础模型，实现了超长序列的高精度、实时单目三维重建与位姿估计。
π0.7：具有涌现能力的可引导通用机器人基础模型（中文翻译）
April 20, 2026
Physical Intelligence 最新论文 π0.7 中文全文翻译 - 一个5B参数的可引导通用机器人基础模型，展现出组合泛化能力，能够开箱即用地执行复杂灵巧任务，实现零样本跨具身迁移。
π0.7：具有涌现能力的可引导通用机器人基础模型（中文全译）
April 20, 2026
本文是Physical Intelligence最新论文π0.7的完整中文翻译。π0.7是一个可引导的通用机器人基础模型，能够在开箱即用情况下执行高度灵巧的长周期任务，实现零样本跨具身迁移，并能通过语言指导学习新任务。
《SONIC: Supersizing Motion Tracking for Natural Humanoid Whole-Body Control》深度解读
April 20, 2026
NVIDIA 研究员将缩放定律（Scaling Law）应用到人形机器人全身控制，通过将模型从 1.2M 放大到 42M 参数、数据集扩大到 1 亿帧（700 小时），得到一个通用人形机器人基础控制器，支持多种输入接口（VR遥操作、视频、VLA模型）。
π0.7: a Steerable Generalist Robotic Foundation Model with Emergent Capabilities 深度解读
April 18, 2026
# 《π0.7: a Steerable Generalist Robotic Foundation Model with Emergent Capabilities》深度解读 **论文信息** - 作者：Bo Ai, Ali Amin, ..., Sergey Levine 等（Physical Intelligence 团队） - 机构：Physical Intelligence - 发表时...
论文解读：Look Before Acting - 增强视觉基础表示的视觉-语言-动作模型
April 16, 2026
本文解读了arXiv 2603.15618论文，该论文揭示了VLA模型中深层视觉敏感性下降的关键问题，并提出DeepVision-VLA框架，在模拟和真实任务上分别超出SOTA 9.0%和7.5%。
From Evaluation to Closed-Loop Improvement: How Community Feedback Makes Robots Smarter
March 26, 2026
When we build a robot or train a new model, we face a fundamental question: **how do we verify that it actually works safely and reliably across all the conditions it might encounter?** In traditional software development, we write unit tests, integration tests, and end-to-end tests. But robotics is...
XRollout Philosophy: The Art of Deliberate Practice
March 26, 2026
> *"Autobots, Roll Out!"* — Optimus Prime The name "XRollout" carries dual meaning. It honors the iconic rallying cry of Optimus Prime from Transformers—a call to action, transformation, and the relentless pursuit of excellence. But more profoundly, **Rollout** represents the cornerstone of our data...
Memory for Robotics: Enhancing Temporal Decision-Making
March 26, 2026
This article breaks down the **MEM (Multi-scale Embodied Memory)** approach from the Physical Intelligence (PI) research project. MEM enables robots to handle long-horizon tasks (up to 15 minutes) by maintaining a structured memory of what they've done, what's still left to do, and where objects are...
Why Language: A Human Brain Perspective on VLA
March 26, 2026
Starting from the human brain. Human conscious thinking primarily consists of five basic functions: understanding, decision-making, recollection, memory, and inhibition. These functions work together to enable planning, problem-solving, communication, and task completion. Many people believe that co...
Community Credit System: The Duolingo Approach to Collaborative Robotics
March 26, 2026
In open-source robotics, we face a classic chicken-and-egg problem: 1. **We need more contributors** to collect data, fix bugs, write documentation, and share knowledge 2. **But contributors need resources** to build their projects — hardware access, GPU time, storage 3. **New participants can't get...
What We Do
March 26, 2026
XRollout was born from a simple but powerful belief: **robotics should be open, accessible, and community-driven**. We believe that the future of robot intelligence shouldn't be locked behind closed doors in corporate labs—it should be built by a global community of hackers, researchers, and enthusi...
Memory for Robotics: Enhancing Temporal Decision-Making
March 26, 2026
Breakdown of Multi-scale Embodied Memory (MEM) - enabling robots to handle long-horizon tasks by remembering what they've done.
Community Credit System: The Duolingo Approach
March 26, 2026
Incentivizing contribution with a credit system where you earn by contributing and redeem for platform resources.
From Evaluation to Closed-Loop Improvement
March 26, 2026
How community feedback closes the data loop and makes robots smarter in real-world scenarios.
Why Language: A Human Brain Perspective on VLA
March 26, 2026
From prefrontal cortex vs basal ganglia to why both vision and language are indispensable in VLA.
XRollout Philosophy: The Art of Deliberate Practice
March 26, 2026
Our complete learning philosophy — four pillars of robot acquisition and the hierarchical data pyramid.
What We Do
March 26, 2026
Our mission - why we started XRollout and what we believe. Robotics should be open, accessible, and community-driven.
SLAM: The Original Memory Theory
March 26, 2026
Simultaneous Localization and Mapping isn't just a robotics algorithm—it's a deep meditation on what memory actually is. When you build SLAM, you're actually building a memory system. SLAM stands for **Simultaneous Localization and Mapping**. But if we translate this into plain language: **"Where am...
Memory as a Service: The Core Problem Revealed by π's MEM
March 26, 2026
Physical Intelligence's π project recently introduced **MEM** (Memory-based Manipulation), bringing memory architectures to the forefront of robot learning. MEM has two key innovations: - **Short-term:** An efficient video encoder based on frame-level π representations for compact recent history - *...
CNN-SLAM vs 传统ORB-SLAM：基于深度学习的SLAM方案真的更好吗？
March 25, 2026
本文在公开TUM数据集上实证对比了CNN-SLAM (DINOv2+LoFTR) vs 传统ORB-SLAM，详细分析了两种方案的优缺点和适用场景。
Experiments (2)
ORB-SLAM (纯Python实现) - TUM fr1_xyz 数据集测试
March 25, 2026
Camera Calibration Pipeline Test - Synthetic Chessboard
March 25, 2026