SEARCH / ARCHIVE全站检索 / 知识库索引

Search Archive全站检索

Search across the atlas, technical notes, projects, and papers.统一检索知识图谱、技术文章、项目与论文笔记。

01 / ATLAS

ATLAS

20

Concepts and the prerequisites between them.知识核心概念及其相互依赖关系。

Mathematics & Control Theory

Mathematical foundations for modeling continuous dynamics, linear spaces, uncertainty, and optimal feedback systems.

Foundations
Computer Science & Software

Core abstractions of computation, algorithmic efficiency, operating kernels, distributed communication, and software systems.

Foundations
Electronics, Embedded & Computing Hardware

Physical sensing, embedded real-time execution, and parallel heterogeneous compute architectures bridging algorithms and hardware.

Foundations
Robotics Fundamentals

Analytical mechanics, spatial transformations, multi-body dynamics, and system middleware for physical robots.

Foundations
Machine Learning & Learning Paradigms

Statistical learning principles for extracting structure, generalizable representations, and adaptive models from data.

Artificial Intelligence
Deep Learning & Neural Architectures

Differentiable neural architectures designed to capture spatial, sequential, relational, and global contextual structure.

Artificial Intelligence
Computer Vision & Spatial Understanding

Computational models that infer semantic entities, temporal motion, and 3D geometric structure from visual observations.

Artificial Intelligence
Language & Multimodal AI

Foundation models for natural language reasoning, cross-modal alignment, and unified multimodal representation.

Artificial Intelligence
AI Agents

Autonomous decision-making systems that couple foundation-model reasoning with memory, tool execution, and environment feedback.

Artificial Intelligence
Robot Manipulation & Dexterity

Physical interaction with objects across single-arm grasping, bimanual coordination, multi-finger hands, and compliant force control.

Robotics
Robot Locomotion

Dynamic balance and terrain traversal for wheeled and legged robots through optimal control and reinforcement learning.

Robotics
Vision-Language Navigation (VLN)

Embodied spatial navigation that grounds natural language instructions into egocentric visual observations and 3D map exploration.

Robotics
Robot Learning

Acquiring closed-loop motor skills through interactive reinforcement learning, sim-to-real transfer, and expert imitation.

Robotics
Generative Modeling & Action Generation

Expressive generative parameterizations for high-dimensional, multimodal continuous robot action trajectories.

Robotics
VLA · Vision-Language-Action Models

Generalist embodied foundation models that map visual observations and natural language instructions directly or hierarchically to robot actions.

Embodied Intelligence
World Models

Predictive simulators of environment physics and future state transitions conditioned on historical observations and candidate actions.

Embodied Intelligence
WAM · World-Action Models

Unified decision architectures that jointly or sequentially couple future world imagination with executable physical action generation.

Embodied Intelligence
Robotic Agents

Long-horizon embodied task orchestration combining high-level task decomposition, skill primitives, closed-loop reflection, and multi-robot coordination.

Embodied Intelligence
Embodied Data & Evaluation

Large-scale data collection paradigms, cross-embodiment state-action interfaces, and reproducible evaluation benchmarks for embodied AI.

Embodied Intelligence
Whole-Body Motion & Control

Coordinated whole-body motion generation, dynamic tracking, and loco-manipulation control for high-DoF humanoid and mobile-manipulator robots.

Embodied Intelligence
02 / BLOG

BLOG

06

Explanations, notes, and working lessons.概念拆解、推导笔记与工程经验。

VLA 架构解析:告别“黑箱”与“空谈”,探寻自动驾驶的“大脑”进化史

从 VA 端到端黑箱、快慢双核 VLM+VA 过渡形态到全可微 VLA 统一大脑,系统拆解视觉(SigLIP+DINOv2 / 3DGS)、语言(LLaMA / MindGPT)、动作解码(Diffusion Transformer)与四阶段演进路线。

embodied-ai · Embodied AI · 25 min
Neural Networks Are Already Clean

The cause of polysemanticity in early model organisms wasn’t superposition, but several ordinary, mostly fixable issues. A working hypothesis from Goodfire on neurons, generalization, and circuits.

deep-learning · Mechanistic Interpretability · 65 min
Thinking About a Bias for Better Semantic Coherence and Reuse

Why current generative next-token prediction and subtractive latent SSL struggle to bias neural networks toward unified, reusable semantic invariances across levels of abstraction.

deep-learning · Representation Learning · 30 min
Constructive Self-Supervised Learning (Part 1): Designing Generalisable Deep Self-Supervision

Designing generalisable deep self-supervision by predicting across an abstraction hierarchy rather than dispersing lower-level details—introducing cI-JEPA (+6.02% linear-probe top-1 over I-JEPA on ImageNet-100).

deep-learning · Representation Learning · 20 min
两万字学习 π0.7:从多模态输入到实时动作生成

结合推理时间线与训练前向数据流,拆解 π0.7 如何将多视角图像、视觉子目标、文本提示与本体感知编码为连续与离散联合动作控制。

embodied-ai · Embodied AI · 45 min
VLA 真的通用吗?—— 机器人具身策略的泛化性与记忆

结合斯坦福 MSL 实验室 Mac Schwager 团队多项实证研究,剖析当前 VLA 与扩散策略的“锁定效应”、动作片段查表记忆化、稀疏自编码器诊断与元认知破局路径。

embodied-ai · Embodied AI · 25 min
03 / PROJECTS

PROJECTS

04

Open systems with evidence attached.可复现、带源码的真实系统实践。

04 / PAPERS

PAPERS

32

Research read as a living trail.学术前沿脉络与方法演进。

ACT: Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware

利用 Transformer、动作分块与时间集成配合 CVAE 潜变量,实现低成本双臂机器人的精细接触式模仿学习。

RSS · 2023 · robot-learning-control
Towards Agentic Robots: Model, Data, Environment, and Harness

系统论述物理世界自主智能体所需的四柱一体工程体系:基础模型、多模态经验数据、交互环境与 Harness 调度。

arXiv · 2026 · robotic-agents
AMO: Adaptive Motion Optimization for Hyper-Dexterous Humanoid Whole-Body Control

融合自适应轨迹优化与强化学习构建高机动人形机器人全身控制框架,协调大跨度俯仰蹲起与双臂操作。

arXiv · 2025 · robot-learning-control
BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Encoders

设计轻量级可训练查询接口 Q-Former,以三目标多掩码两阶段训练高效连接冻结视觉编码器与冻结大语言模型。

ICML / arXiv · 2023 · vlm-foundations
CLONE: Closed-Loop Whole-Body Humanoid Teleoperation for Long-Horizon Tasks

引入视觉闭环反馈克服纯本体里程计漂移的人形机器人全身遥操作控制框架,服务长时程移动操作任务。

arXiv · 2025 · robot-learning-control
DAgger: A Reduction of Imitation Learning to No-Regret Online Learning

通过专家在策略实际访问状态分布上迭代纠偏标注,从理论上破解模仿学习的复合误差漂移难题。

AISTATS / arXiv · 2011 · robot-learning-control
DETR: End-to-End Object Detection with Transformers

使用可学习 Object Query 与二分图匹配集合损失取代锚框与 NMS,确立查询式视觉感知底座。

ECCV / arXiv · 2020 · vlm-foundations
Diffusion Policy: Visuomotor Policy Learning via Action Diffusion

将机器人视觉运动策略建模为动作空间上的条件去噪扩散过程,从原理上统一多峰动作表示与滚动控制。

IJRR / RSS · 2024 · robot-learning-control
DINOv2: Learning Robust Visual Features without Supervision

在精选 1.42 亿无标注数据 LVD-142M 上规模化自蒸馏 ViT,为开放世界密集定位与机器人视觉策略提供通用骨干。

TMLR / arXiv · 2024 · vlm-foundations
DPO: Direct Preference Optimization

将带 KL 约束的 RLHF 奖励最大化目标解析化简为基于偏好对比的二元交叉熵分类损失,奠定高效策略对齐基础。

NeurIPS · 2023 · robotic-agents
Flamingo: A Visual Language Model for Few-Shot Learning

借助 Perceiver Resampler 压缩变长图像/视频特征并接入冻结大语言模型门控交叉注意力层,打通少样本视觉语言交互。

NeurIPS · 2022 · vlm-foundations
GPT-Policy: In-Context Robot Learning with VLM Agents

在不更新任何模型权重的前提下,通过组织五类运行时视觉-轨迹上下文驱动通用 VLM 智能体适应真实机器人操作。

arXiv · 2026 · robotic-agents
HAMT: History Aware Multimodal Transformer for Vision-and-Language Navigation

用层级 Transformer 显式编码完整历史全景观测与朝向序列,实现跨长轨迹时序因果对齐。

arXiv / NeurIPS · 2021 · vision-language-navigation
HarnessVLN: Unifying Training-Free Embodied Navigation through an Agent Harness

面向多模态大模型零训练导航的 Agent Harness,以分层事件记忆与时空图 ST Graph 约束规划并校验终止条件。

arXiv · 2026 · vision-language-navigation
HoloAgent-0: A Unified Embodied Agent Framework with 3D Spatial Memory

为具身多模态智能体引入统一三维空间记忆表征,统筹长程导航感知推理与跨区域物理执行。

arXiv · 2026 · robotic-agents
LLaVA: Visual Instruction Tuning

用极简线性投影连接 CLIP 视觉编码器与 LLaMA,以多模态指令微调确立现代 MLLM/VLM 极简通用架构。

NeurIPS / arXiv · 2023 · vlm-foundations
MEM: Multi-Scale Embodied Memory for Vision Language Action Models

为 VLA 引入高频动作块与长时程跨幕结构化记忆检索的多尺度记忆架构,突破短视窗口瓶颈。

arXiv · 2026 · vla
深度解析:理想 MindVLA 和小米 ORION 的 VLA 核心技术栈与具身智能的未来

解剖理想 MindVLA(3DGS 空间智能 / 从零预训练 MindGPT / Diffusion + ODE Sampler)与小米 ORION(QT-Former 长时序压缩 / 可微规划 Token),并梳理数据炼金术、RLHF、世界模型闭环与四大落地挑战。

Technical Deep Dive · ORION (arXiv:2503.19755) · 2025 · autonomous-driving-vla
OpenVLA: An Open-Source Vision-Language-Action Model

基于开源多模态骨干与 Open X-Embodiment 970k 轨迹训练的 7B 开源视觉-语言-动作基础模型。

arXiv · 2024 · vla
Pathdreamer: A World Model for Indoor Navigation

结合 3D 语义点云重投影与生成式补全合成未走入房间的高分辨率全景未来观测,辅助导航前瞻规划。

ACL ALVR · 2021 · world-action-models
π₀: A Vision-Language-Action Flow Model for General Robot Control

在预训练 VLM 上引入独立 Action Expert 与 Flow Matching 连续动作块生成头,支撑 50Hz 灵巧双臂控制。

arXiv · 2024 · vla
π₀.₅: A Vision-Language-Action Model with Open-World Generalization

结合跨场景异构协同训练与高层语义子任务预测,将 VLA 泛化至训练未见过的真实家庭环境。

arXiv · 2025 · vla
PREVALENT: Towards Learning a Generic Agent for Vision-and-Language Navigation via Pre-Training

通过视听图文动作联合预训练为 VLN 构建通用视觉-语言表征底座,摆脱单一任务从头训练限制。

CVPR · 2020 · vision-language-navigation
R2R: Vision-and-Language Navigation in Real Environments

基于 Matterport3D 实景扫描提出 Room-to-Room 基准与全景导航图,确立 VLN 标准任务范式。

CVPR · 2018 · vision-language-navigation
RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control

将互联网图文预训练 VLM 直接微调为输出离散动作词元的具身策略,提出 VLA 范式。

arXiv · 2023 · vla
SimCLR: A Simple Framework for Contrastive Learning of Visual Representations

系统化证明数据增强组合、非线性投影头与归一化温度对比损失能够学到堪比有监督预训练的视觉表征。

ICML / arXiv · 2020 · vlm-foundations
Tactical Rewind: Self-Correction via Backtracking in Vision-and-Language Navigation

让导航智能体在察觉偏航或指令对齐度下降时,主动沿访问路径回溯重选分支,缓解不可逆错误。

CVPR · 2019 · vision-language-navigation
Thea: Towards the Harness of Embodied Agents

将软件工程中的 Agent Harness 迁移至物理机器人,以持久化场景图与评估退出码闭合感知—执行—重试环路。

Project / Preprint · 2026 · robotic-agents
Video Generators are Robot Policies

探索以预训练视频生成模型为核心统一未来物理帧演进预测与可执行机器人动作轨迹推导。

arXiv · 2025 · world-action-models
Reasoning via Vision-Language World Models (VLWM)

利用分层视频字幕树与 LLM 自优化构建语言空间抽象世界模型,并以自监督代价评判器支持 System 2 规划搜索。

arXiv · 2025 · world-action-models
World-Action Models for Robot Learning and Control: A Survey

系统拆解 WAM 五大组成维度与两大系统范式(端到端 vs 双模块),警示切勿仅凭 FVD/LPIPS 评判控制效用。

arXiv · 2026 · world-action-models
World Action Models: The Next Frontier in Embodied AI

OpenMOSS 团队首发 WAM 系统化定义与架构分类体系,形式化未来状态预测与动作生成的联合概率分布。

arXiv · 2026 · world-action-models