Highly Dynamic Quadruped Locomotion via Whole-Body Impulse Control and Model Predictive Control MIT Biomimetics Lab(D. Kim, J. Di Carlo, B. Katz, G. Bledt, S. Kim)· Science Robotics 2019 MIT Cheetah 3 用"凸 MPC + 整机冲量控制"实现盲行走、越障与抗扰,是"模型类足式控制"的教科书。arXiv:1909.06586 →
Learning agile and dynamic motor skills for legged robots ETH Zurich / MPI(J. Hwangbo 等)· Science Robotics 2019 深度 RL 大规模落地真实四足的里程碑:ANYmal 学会跳跃、爬坡等敏捷技能。arXiv:1901.08652 →
Learning quadrupedal locomotion over challenging terrain ETH Zurich(J. Lee 等)· Science Robotics 2020 RL + 感知让 MIT Mini Cheetah 在复杂地形自主奔跑,理解"仿真训练 + 实物部署"全流程。arXiv:2010.11251 →
Learning to Walk in Minutes Using Massively Parallel Deep Reinforcement Learning ETH Zurich(N. Rudin 等)· CoRL 2021 用 Isaac Gym 在 GPU 上并行数千个仿真环境,几分钟训出四足步态——现在 legged-gym 生态的源头。arXiv:2109.11978 →
Sim-to-Real Transfer for Biped Locomotion UBC / BYU(Z. Xie 等)· IROS 2019 双足机器人 Cassie 上 RL 从仿真迁移到实物,学会平衡与行走,理解"域随机化"。arXiv:1903.01390 →
DeepMimic: Example-Guided Deep Reinforcement Learning of Physics-Based Character Skills UC Berkeley(X. B. Peng 等)· SIGGRAPH 2018 用动捕参考动作引导 RL 学复杂技能,"动作模仿"的经典范式。arXiv:1808.00101 →
② 人形整机控制:WBC 与 MPC
Optimization-based locomotion planning, estimation, and control design for the atlas humanoid robot MIT CSAIL(S. Kuindersma 等)· IJRR 2016 Atlas 基于优化的整机控制体系:轨迹规划、状态估计、MPC、分层 WBC 一次讲全,体系感最强的一篇。arXiv:1507.03552 →
Whole-Body Dynamic Behavior and Control of Human-like Robots Stanford University(L. Sentis, O. Khatib)· IJRR 2006 分层整机控制(WBC)的奠基之作:任务层级 + 零空间投影,理解"优先级"思想必读。SAGE 原文 →
A Direct Method for Trajectory Optimization of Rigid Bodies Through Contact MIT(M. Posa, C. Cantu, R. Tedrake)· IJRR 2014 把"接触"直接纳入轨迹优化的经典方法,现代接触丰富 MPC 的理论地基。SAGE 原文 →
Whole Body Model Predictive Control with a Memory of Motion: Experiments on a Torque-Controlled Talos LAAS-CNRS / Inria(E. Dantec 等)· ICRA 2021 用"动作记忆库"热启动整机 MPC,让 Talos 人形实时完成复杂操作,整机 MPC 工程化的代表。HAL:hal-02995796 →
Solving Rubik's Cube with a Robot Hand OpenAI· 2019 仿真 RL + 域随机化 + 自动域适应(ADA),让机械手盲拧魔方,理解"仿真到实物"的最强案例。arXiv:1910.07113 →
Learning Complex Dexterous Manipulation with Deep Reinforcement Learning and Demonstrations UC Berkeley / Google Brain(A. Rajeswaran 等)· RSS 2018 RL + 示范数据结合学习灵巧手复杂操作,演示引导(DAPG)的经典出处。arXiv:1709.10087 →
AnyGrasp: Robust and Efficient Grasp Perception in Spatial and Temporal Domains 清华大学 / 腾讯(H. Fang 等)· IEEE T-RO 2023 从点云实时估计抓取位姿,对未见物体泛化强,抓取感知落地的常用工具(有开源推理代码)。arXiv:2212.08333 →
DexGraspNet: A Large-Scale Robotic Dexterous Grasp Dataset for General Objects Based on Simulation 北京大学 PKU-EPIC· NeurIPS 2022 大规模灵巧抓取数据集 + 配套生成方法(详见第 4 节数据集表),训练灵巧手抓取的起点。arXiv:2210.02697 →
④ VLA 与具身智能
RT-1: Robotics Transformer for Real-World Control at Scale Google Robotics(A. Brohan 等)· RSS 2023 13 台机器人、130k 条示教轨迹训练出的 Transformer 策略,"机器人 Transformer"路线开山之作。arXiv:2212.06817 →
RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control Google DeepMind(A. Brohan 等)· CoRL 2023 把视觉语言模型变成视觉语言动作模型,让"网页知识"迁移到机器人控制。arXiv:2307.15818 →
PaLM-E: An Embodied Multimodal Language Model Google / MPI(D. Driess 等)· ICML 2023 5620 亿参数的具身多模态大模型,"大模型 + 机器人"路线必读。arXiv:2303.03378 →
Open X-Embodiment: Robotic Learning Datasets and RT-X Models Open X-Embodiment 协作组(多机构)· ICRA 2024 22 种机器人形态、百万级轨迹的跨形态数据集 + RT-X 模型,具身智能的"数据底座"。arXiv:2310.08864 →
OpenVLA: An Open-Source Vision-Language-Action Model Stanford / Google DeepMind 等(M. J. Kim 等)· CoRL 2024 7B 开源 VLA 模型,权重与训练代码全开放,是入门 VLA 微调的首选。arXiv:2406.09246 →
Octo: An Open-Source Generalist Robot Policy UC Berkeley / Stanford 等(Octo Model Team)· RSS 2024 开源通用机器人策略 Transformer,支持跨形态、可微调,工程化程度高。arXiv:2405.12213 →
⑤ 世界模型
Recurrent World Models Facilitate Policy Evolution Uber AI Labs(D. Ha, J. Schmidhuber)· NeurIPS 2018 "世界模型"概念三件套(Vision/记忆/Controller),在赛车游戏里展示"在想象中学习"的威力。arXiv:1809.01999 →
Dream to Control: Learning Behaviors by Latent Imagination DeepMind / Google Brain(D. Hafner 等)· ICLR 2020 Dreamer:在潜在空间"做梦"训练策略,世界模型用于控制的开端。arXiv:1912.01603 →
Mastering Atari with Discrete World Models DeepMind(D. Hafner 等)· ICLR 2021 DreamerV2 用离散潜变量世界模型登顶 Atari,方法干净利落,论文写作也是范本。arXiv:2010.02193 →
Mastering Diverse Domains through World Models DeepMind(D. Hafner 等)· 2023 DreamerV3:一套超参数通吃 150+ 任务,世界模型通用化的代表作。arXiv:2301.04104 →
Genie: Generative Interactive Environments Google DeepMind(J. Bruce 等)· 2024 从互联网视频直接生成可交互 2D 环境,"生成式世界模型"的代表作,值得追踪其后继(Genie 2 等)。arXiv:2402.15391 →
Sensorless Field Oriented Control of a PMSM(AN1078 应用笔记) Microchip· 官方应用笔记 PMSM 无传感器 FOC 的工程实现指南:滑模观测器、启动流程、调参步骤,单片机实现 FOC 的必读文档。Microchip AN1078 →
Field Oriented Control of 3-Phase AC-Motors(TI BPRA073 应用报告) Texas Instruments· 经典应用报告 FOC 数学原理与矢量控制的经典入门文档(有中文译本流传),把坐标系变换讲得最清楚。TI BPRA073(PDF)→
Proprioceptive Actuator Design in the MIT Cheetah: Impact Mitigation and High-Bandwidth Physical Interaction for Dynamic Legged Robots MIT(P. M. Wensing 等)· IEEE T-RO 2017 "本体感觉驱动器"设计:无传感器力矩估计、抗冲击、高带宽力控——理解机器人关节"电机+减速器+编码器"为何这样设计。MIT Biomimetics 页面 →
VESC 开源 ESC 项目(含 Benjamin Vedder 硕士论文《VESC – Open Source ESC》) B. Vedder· KTH 硕士论文 2015 + 持续开源维护 从 FOC 理论到"能跑的电机控制器"的完整工程范本,固件全程开源,是 ODrive/SimpleFOC 之外的另一极。VESC 固件 GitHub →