Atlas / L1 · Foundations
Robotics Fundamentals.机器人学基础.
Analytical mechanics, spatial transformations, multi-body dynamics, and system middleware for physical robots.描述机器人机械本体、空间刚体变换、多体动力学建模与系统软件通信的经典机器人学基础。
F05 · L1 · Structured Network · 4 Pillars
ATLAS知识图谱 / ATLAS
Mathematics & Control TheoryMathematical foundations for modeling continuous dynamics, linear spaces, uncertainty, and optimal feedback systems.FoundationsComputer Science & SoftwareCore abstractions of computation, algorithmic efficiency, operating kernels, distributed communication, and software systems.FoundationsElectronics, Embedded & Computing HardwarePhysical sensing, embedded real-time execution, and parallel heterogeneous compute architectures bridging algorithms and hardware.FoundationsRobotics FundamentalsAnalytical mechanics, spatial transformations, multi-body dynamics, and system middleware for physical robots.FoundationsMachine Learning & Learning ParadigmsStatistical learning principles for extracting structure, generalizable representations, and adaptive models from data.Artificial IntelligenceDeep Learning & Neural ArchitecturesDifferentiable neural architectures designed to capture spatial, sequential, relational, and global contextual structure.Artificial IntelligenceComputer Vision & Spatial UnderstandingComputational models that infer semantic entities, temporal motion, and 3D geometric structure from visual observations.Artificial IntelligenceLanguage & Multimodal AIFoundation models for natural language reasoning, cross-modal alignment, and unified multimodal representation.Artificial IntelligenceAI AgentsAutonomous decision-making systems that couple foundation-model reasoning with memory, tool execution, and environment feedback.Artificial IntelligenceRobot Manipulation & DexterityPhysical interaction with objects across single-arm grasping, bimanual coordination, multi-finger hands, and compliant force control.RoboticsRobot LocomotionDynamic balance and terrain traversal for wheeled and legged robots through optimal control and reinforcement learning.RoboticsVision-Language Navigation (VLN)Embodied spatial navigation that grounds natural language instructions into egocentric visual observations and 3D map exploration.RoboticsRobot LearningAcquiring closed-loop motor skills through interactive reinforcement learning, sim-to-real transfer, and expert imitation.RoboticsGenerative Modeling & Action GenerationExpressive generative parameterizations for high-dimensional, multimodal continuous robot action trajectories.RoboticsVLA · Vision-Language-Action ModelsGeneralist embodied foundation models that map visual observations and natural language instructions directly or hierarchically to robot actions.Embodied IntelligenceWorld ModelsPredictive simulators of environment physics and future state transitions conditioned on historical observations and candidate actions.Embodied IntelligenceWAM · World-Action ModelsUnified decision architectures that jointly or sequentially couple future world imagination with executable physical action generation.Embodied IntelligenceRobotic AgentsLong-horizon embodied task orchestration combining high-level task decomposition, skill primitives, closed-loop reflection, and multi-robot coordination.Embodied IntelligenceEmbodied Data & EvaluationLarge-scale data collection paradigms, cross-embodiment state-action interfaces, and reproducible evaluation benchmarks for embodied AI.Embodied IntelligenceWhole-Body Motion & ControlCoordinated whole-body motion generation, dynamic tracking, and loco-manipulation control for high-DoF humanoid and mobile-manipulator robots.Embodied IntelligenceBLOG技术文章 / BLOG
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 minNeural Networks Are Already CleanThe 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 minThinking About a Bias for Better Semantic Coherence and ReuseWhy 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 minConstructive Self-Supervised Learning (Part 1): Designing Generalisable Deep Self-SupervisionDesigning 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 minVLA 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 minPAPERS学术论文 / PAPERS
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-controlTowards 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-agentsAMO: 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-controlBLIP-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-foundationsCLONE: 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-controlDAgger: 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-controlDETR: 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-foundationsDiffusion 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-controlDINOv2: 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-foundationsDPO: 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-agentsFlamingo: 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-foundationsGPT-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-agentsHAMT: 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-navigationHarnessVLN: 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-navigationHoloAgent-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-agentsLLaVA: 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-foundationsMEM: 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-vlaOpenVLA: 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 · vlaPathdreamer: 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-modelsπ₀: 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 · vlaπ₀.₅: 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 · vlaPREVALENT: 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-navigationR2R: 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-navigationRT-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 · vlaSimCLR: 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-foundationsTactical 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-navigationThea: 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-agentsVideo 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-modelsReasoning 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-modelsWorld-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-modelsWorld Action Models: The Next Frontier in Embodied AIOpenMOSS 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-modelsPROJECTS项目案例 / PROJECTS
LoneRoverA complete ROS 2 navigation stack for an Ackermann vehicle, connecting state estimation, planning, control, and safety on a real competition platform.Real-world robotics · ROS 2 · EKF · Nav2 · MPPI · RPPRealMan AgentAn embodied ReAct agent framework with Working Memory and SQLite Episodic Memory for long-horizon mobile manipulation, connecting navigation, YOLO-seg + FoundationPose 6D pose estimation, and grasping on RealMan hardware.Embodied agent · ReAct · Episodic Memory · FoundationPose · RealMancorrMOTRA MOTR-oriented research system that separates geometry/motion, appearance identity, and existence/lifecycle into explicit state domains.Research prototype · MOTR · DETR · TrackPool · StateTSMA Temporal Shift Module pipeline for real-time gesture recognition, with data preparation, training configuration, checkpoint loading, and webcam inference.Temporal vision / HRI · TSM · Video · HRI · PyTorchNo matching records.未找到匹配的记录。
Atlas / L1 · Foundations
Analytical mechanics, spatial transformations, multi-body dynamics, and system middleware for physical robots.描述机器人机械本体、空间刚体变换、多体动力学建模与系统软件通信的经典机器人学基础。
F05 · L1 · Structured Network · 4 Pillars