04 / PAPERS04 / 论文脉络
Papers & Reading Notes论文与精读笔记
Research papers, technical analysis, and comparisons.围绕具身智能的研究论文、技术分析与方法对比。
32 PAPERS / 06 RESEARCH AREAS · ZOTERO + PKA篇文献 / 06 个研究方向 · ZOTERO + PKA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。
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。
RESEARCH INDEX / 06 EMBODIED DIRECTIONS研究方向索引 / 六大具身智能脉络
Browse literature by research trajectory按具身智能研究主线浏览关键论文
Grounded in local Zotero collections and Personal Knowledge Atlas (PKA) reading archives.基于本地 Zotero 题录与 Personal Knowledge Atlas(PKA)精读档案整理,点击左侧领域切换脉络。
ACTIVE DIRECTION当前方向
Vision-Language-Action (VLA)视觉-语言-动作模型
End-to-end generalist robot policies unifying vision-language backbones, continuous flow/diffusion action heads, and embodied memory.连接多模态骨干、连续动作生成头(Flow Matching / Action Token)与长程具身记忆的通用操作策略。
ACTIVE DIRECTION当前方向
Vision-Language Navigation (VLN)视觉语言导航
Instruction-following and object-goal navigation across discrete nav-graphs, history-aware transformers, and agent harness execution.覆盖离散导航图基准、历史时空建模,到零样本智能体调度器闭环的具身导航研究。
ACTIVE DIRECTION当前方向
World Models & WAM世界模型与世界动作模型
Predictive future-state modeling—from visual indoor panoramas and video policy generators to language-space world models and coupled World-Action Models.将未来世界状态预测(视频、视觉语言状态树与隐空间推演)与可执行动作生成耦合的研究分支。
ACTIVE DIRECTION当前方向
Robot Learning & Control机器人学习与控制
Online imitation error correction, action chunking, diffusion visuomotor policies, and adaptive humanoid whole-body control.探讨复合误差缓解、动作分块、扩散多峰动作分布,以及面向双臂与人形机器人的全身运动控制。
ACTIVE DIRECTION当前方向
LLMs & Embodied Agents大模型与具身智能体
Preference alignment, 3D spatial memory, tool/scene-graph agent harnesses, and in-context robot policy adaptation.围绕偏好对齐、三维空间记忆、物理世界 Agent Harness 运行时设施与免参数更新上下文策略。
ACTIVE DIRECTION当前方向
Vision-Language & Representations视觉语言与表征基础
Self-supervised visual representations, set-based object queries, and frozen vision-to-LLM alignment architectures.构筑具身策略感知底座的自监督表征、无锚框集合预测及跨模态图文对齐接口。
READING TRAIL / CHRONOLOGICAL VIEW研究脉络 / 按发表年份排序
VLA Generalist PoliciesVLA 策略演进脉络
From discrete action tokenization in RT-2 to open-weights OpenVLA, Flow Matching in π₀/π₀.₅, and multi-scale memory in MEM.从 RT-2 离散动作词元、OpenVLA 开源权重到 π₀/π₀.₅ 连续流匹配与 MEM 多尺度具身记忆的演进主线。
Time orders the readings. It does not imply method inheritance.按论文发表时间排序便于对照演进,不代表单一的线性方法继承关系。
Vision-Language-Action (VLA) / CHRONOLOGICAL LEDGER视觉-语言-动作模型 / 文献归档清单
05 RECORDS- 2023
PKA / 10 来源研究 / Robot Learning · VLA(SayCan、RT系列与RT-2演进)
- 2024
Zotero / VLA · PKA / 领域索引 / VLA
- 2024
Zotero / VLA · PKA / [论文精读] π₀与π₀-FAST
- 2025
Zotero / VLA · PKA / 一文梳理 π 系列 VLA 模型技术演进
- 2026
Zotero / Embodied Agents & VLA
READING TRAIL / CHRONOLOGICAL VIEW研究脉络 / 按发表年份排序
Autonomous Driving VLA & Physical Agents自动驾驶 VLA 架构演进脉络
End-to-end physical AI architectures uniting 3DGS perception, from-scratch MindGPT / VLM reasoning, Diffusion/VAE trajectory generation, and World Model + RLHF closed-loop alignment (MindVLA & ORION).拆解理想 MindVLA(3DGS / MindGPT / Diffusion)与小米 ORION(QT-Former / 规划 Token)核心技术栈、世界模型与 RLHF 闭环及物理智能体落地挑战。
Time orders the readings. It does not imply method inheritance.按论文发表时间排序便于对照演进,不代表单一的线性方法继承关系。
Vision-Language-Action (VLA) / CHRONOLOGICAL LEDGER视觉-语言-动作模型 / 文献归档清单
01 RECORDS- 2025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Technical Deep Dive · ORION (arXiv:2503.19755)
公众号「青稞具身智能」· 知乎专栏 @我要吃鸡腿
READING TRAIL / CHRONOLOGICAL VIEW研究脉络 / 按发表年份排序
Vision-Language Navigation视觉语言导航脉络
Tracing R2R task formulation, backtracking self-correction, multimodal pre-training, long-horizon history encoding, and harness verification.梳理 R2R 基准建立、Tactical Rewind 回溯纠错、PREVALENT 预训练、HAMT 历史编码到 HarnessVLN 调度闭环。
Time orders the readings. It does not imply method inheritance.按论文发表时间排序便于对照演进,不代表单一的线性方法继承关系。
Vision-Language Navigation (VLN) / CHRONOLOGICAL LEDGER视觉语言导航 / 文献归档清单
05 RECORDS- 2018
Zotero / VLN · PKA / VLN 最具影响力研究梳理
- 2019
- 2020
Zotero / VLN · PKA / VLN 代表性研究盘点
- 2021
Zotero / VLN · PKA / VLN 代表性研究盘点
- 2026
PKA / 10 来源研究 / 领域索引 / VLN(南大开源 HarnessVLN)
READING TRAIL / CHRONOLOGICAL VIEW研究脉络 / 按发表年份排序
World & World-Action Models世界模型与 WAM 脉络
Following indoor visual lookahead (Pathdreamer), video-policy unification, caption-tree reasoning (VLWM), and formal WAM taxonomies.从室内未来视角预测 Pathdreamer、视频策略统一,到语言状态推理 VLWM 与 WAM 系统化范式综述。
Time orders the readings. It does not imply method inheritance.按论文发表时间排序便于对照演进,不代表单一的线性方法继承关系。
World Models & WAM / CHRONOLOGICAL LEDGER世界模型与世界动作模型 / 文献归档清单
05 RECORDS- 2021
Zotero / VLN & World Models
- 2025
PKA / 10 来源研究 / 领域索引 / 世界模型-WAM(Video Generators are Robot Policies)
- 2025
PKA / 10 来源研究 / 领域索引 / 世界模型-WAM(使用视觉语言世界模型进行推理规划)
- 2026
PKA / 10 来源研究 / 领域索引 / 世界模型-WAM
- 2026
PKA / 10 来源研究 / 领域索引 / 世界模型-WAM(世界动作模型WAMs最新综述)
READING TRAIL / CHRONOLOGICAL VIEW研究脉络 / 按发表年份排序
Robot Learning & Whole-Body Control机器人学习与控制脉络
Connecting DAgger covariate-shift mitigation, ACT chunking, Diffusion Policy, and humanoid torso/teleoperation whole-body controllers.从 DAgger 解决状态分布偏移,到 ACT 动作分块、Diffusion Policy 条件去噪及 AMO/CLONE 人形全身控制。
Time orders the readings. It does not imply method inheritance.按论文发表时间排序便于对照演进,不代表单一的线性方法继承关系。
Robot Learning & Control / CHRONOLOGICAL LEDGER机器人学习与控制 / 文献归档清单
05 RECORDS- 2011
Zotero / Robot Learning · PKA(ACT 背景综述引用)
- 2023
Zotero / Robot Learning · PKA / [论文精读] ACT
- 2024
PKA / 10 来源研究 / Robot Learning(万字长文解读 Diffusion Policy)
- 2025
Zotero / Humanoid Whole-Body Control
- 2025
Zotero / Humanoid Whole-Body Control
READING TRAIL / CHRONOLOGICAL VIEW研究脉络 / 按发表年份排序
Embodied Agent Architecture具身智能体系统脉络
Bridging preference optimization (DPO) to unified 3D-memory agents (HoloAgent-0), physical harnesses (Thea), and in-context learning.从 DPO 偏好对齐到三维空间记忆 HoloAgent-0、物理执行环路 Thea Harness 与上下文策略学习。
Time orders the readings. It does not imply method inheritance.按论文发表时间排序便于对照演进,不代表单一的线性方法继承关系。
LLMs & Embodied Agents / CHRONOLOGICAL LEDGER大模型与具身智能体 / 文献归档清单
05 RECORDS- 2023
Zotero / LLM Alignment & Agents
- 2026
Zotero / Embodied Agents
- 2026Thea: Towards the Harness of Embodied AgentsProject / Preprint
PKA / 10 来源研究 / 领域索引 / LLM-Agent(Towards the Harness of Embodied Agents)
- 2026
PKA / 10 来源研究 / LLM-Agent · Zotero / In-Context Robot Learning
- 2026
Zotero / Embodied Agents
READING TRAIL / CHRONOLOGICAL VIEW研究脉络 / 按发表年份排序
Visual & VLM Foundations视觉语言表征脉络
Trace contrastive features (SimCLR, DINOv2) alongside set queries (DETR) and cross-modal bridges (Flamingo, BLIP-2, LLaVA).沿着对比与自监督视觉特征(SimCLR / DINOv2)、集合查询(DETR)与跨模态桥梁(Flamingo / BLIP-2 / LLaVA)阅读。
Time orders the readings. It does not imply method inheritance.按论文发表时间排序便于对照演进,不代表单一的线性方法继承关系。
Vision-Language & Representations / CHRONOLOGICAL LEDGER视觉语言与表征基础 / 文献归档清单
06 RECORDS- 2020
Zotero / Visual Representation & VLM
- 2020
PKA / 10 来源研究 / DETR 从图像到无序对象集合.md
- 2022
Zotero / VLM Foundations
- 2023
Zotero / VLM · PKA / 10 来源研究 / BLIP-2 查询式跨模态接口.md
- 2023LLaVA: Visual Instruction TuningNeurIPS / arXiv
Zotero / VLM · PKA / LLaVA 系列演化综述
- 2024
Zotero / Visual Representation & VLM
RESEARCH TRAIL
No reading trails matched.
Select one of the research directions above to inspect its chronological paper trail.