跨本体与控制 · Cross-Embodiment and Control

近一年 loco-manipulation 研究进展

Progress in Loco-Manipulation

更新于

30 秒结论

先读这个判断

本文讨论 2025 年 7 月 19 日至 2026 年 7 月 19 日公开的 loco-manipulation 研究,即机器人在移动、保持平衡或改变支撑关系的同时完成物体交互。范围同时覆盖足式机器人、轮式移动操作与人形机器人,但重点放在“移动与操作是否真正耦合”,而不是一般导航后再抓取。

精读论文
21
正式证据事件
21
知识单元
EA-MODEL, EA-EVAL, EA-XEMBODIMENT, EA-SENSOR
审计状态
当前结算 run 已通过仓库校验

v0 · 完整正文

研究综述

研究边界

本文讨论 2025 年 7 月 19 日至 2026 年 7 月 19 日公开的 loco-manipulation 研究,即机器人在移动、保持平衡或改变支撑关系的同时完成物体交互。范围同时覆盖足式机器人、轮式移动操作与人形机器人,但重点放在“移动与操作是否真正耦合”,而不是一般导航后再抓取。

这是一份范围综述:检索得到 556 篇候选,恢复了 76 篇完整非 OCR 正文,最终对 21 篇代表性论文完成逐条证据核验。它适合判断技术路线和研究空白,不适合据此给所有方法做统一排行榜,因为各论文的机器人、任务、成功定义和试验次数差异很大。

中心判断

近一年最重要的变化,不是机器人又多会了几个“走过去再抓”的任务,而是研究重心开始从模块拼接转向三种原生耦合:全身动作空间的统一控制、跨来源数据的统一学习,以及视觉—地形—触觉—故障状态的统一反馈。结果是低层全身协调已经在受控场景中形成可复用能力,数据扩展也出现了清晰方法论;但自主性瓶颈随之上移到物体状态估计、长时序记忆、任务规划和真实世界失配。换言之,领域正从“能不能同时走和操作”转向“怎样在未知场景里持续、可靠地做对”。

一、控制范式:下肢不再只是会走路的底盘

过去常见系统把上肢逆解和下肢行走策略分开,再用速度与高度指令粘合。它容易训练,却天然排除了脚踩踏板、深蹲取物、边走边维持受力等动作。近一年的论文开始系统验证这种分解的代价。

OpenHLM 在真实 Unitree G1 上比较了多种遥操作与动作表示。其关键发现不是某个 VLA 主干更强,而是完整暴露全身关节的 joint-based interface,能表达解耦控制或三点 VR 接口做不到的深蹲、脚部操作和全身协同。它说明“策略上限”首先受动作接口约束:数据采集时没让操作者使用腿和腰,后续模型再大也学不到这些动作。

这一趋势在不同形态上有相似结论。MLM 用同一四足臂控制器完成多任务轨迹跟踪,真实任务中遥操作成功率达 98%/100%,而由扩散策略提供轨迹时降至 80%/85%;低层控制已较稳定,高层轨迹质量成为主要误差源。TA-WBC 又把地形编码和相对于接触平面的末端目标引入控制,在“取瓶—上五级台阶—投放—返回”的 staged route 上连续完成 5 次且无跌倒或踉跄。这里真正的新意是操作目标不再绑定波动的机身坐标,而是绑定脚部接触几何。

另一条路线保留低层策略,把灵活性放到在线规划。Sumo 在预训练全身策略的低维命令空间里运行采样式 MPC。Spot 在真实硬件上用手臂、躯干和腿共同竖起 15 kg 轮胎——超过论文给出的 11 kg 手臂额定举升能力——10 次全部成功。它表明“学习稳定运动原语 + 测试时规划”能够处理大而重的物体,并通过修改模型或代价函数迁移任务;代价是仍依赖动捕、物体模型和人工设计任务代价。

因此,控制层的阶段性结论是:原生全身动作空间已显示出明确收益,但“统一策略”不是“端到端解决一切”。更可行的系统往往仍是分层的,只是分层边界从“上肢/下肢”转向“任务意图/全身执行”。

二、数据路线:从昂贵全身遥操作,转向有结构的异构数据组合

人形 loco-manipulation 的数据成本很高:操作员要同时控制步态、姿态、双臂和接触,还要承担硬件复位与安全成本。近一年最扎实的进展之一,是论文不再只说“需要更多数据”,而是开始回答不同数据源究竟提供什么。

HumanoidMimicGen 将一次人体示范拆成技能、重定向并进行全身规划与随机化,在九个仿真任务上把 DexMimicGen+ 的平均策略表现从 0.33 提高到 0.89。它证明结构化生成可以放大稀疏示范,但条件也很清楚:技能分段、先后关系、环境和成功判据仍需人工定义。

真实采集侧,HuMI 用穿戴式追踪和人机回路 IK 预览实现无机器人采集。在 15 分钟比较中,它采集 62 条轨迹,对照系统为 28 条;验收率为 96.7% 对 64.3%,单条合格数据耗时降至对照的 30%。EgoHumanoid 则把 300 条第一视角人类示范与 100 条机器人示范对齐共训:域内得分从 robot-only 的 59% 提至 78%,泛化场景从 31% 提至 82%。这说明人类数据更像“场景与运动先验”,但仍需要机器人数据锚定动作语义,而且导航/接近通常比精细接触更容易迁移。

OpenHLM 对数据来源的分工给出了更直接的实验:四个未被全身遥操作覆盖的任务上,加入同形态、站立不走的操作数据后,平均任务进度由 33% 升至 87%,接近 12 任务全遥操作 oracle 的 94%。Robot-free HuMI 数据能教会新物体和新语言,但在当前规模下不擅长注入训练集中不存在的新动作。这个结果把“异构共训”从口号变成了设计原则:

  1. 用少量全身遥操作建立形态与动作锚点;
  2. 用便宜的同形态静态操作补精细动作;
  3. 用人类或无机器人数据扩展物体、语言与场景;
  4. 用仿真生成覆盖初始状态和布局,而不是期望单一数据源包办所有能力。

WholeBodyVLA 进一步表明,无动作标注的人类视频也能提供统一潜在动作先验;移除该潜在模型后,论文报告成功率下降 38.7 个百分点。不过它和上述结果共同提醒:跨形态数据最先改善的是表示与泛化,不会自动解决接触动力学和机器人可执行性。

三、感知与反馈:视觉之外,地形、触觉和故障开始进入闭环

若全身控制只看本体状态,它很难判断脚下坡度、手中滑移或执行器退化。近一年出现了三类值得重视的反馈通道。

第一类是地形。TA-WBC 的足端多环高程编码让控制器提前调整姿态和落脚点,说明 loco-manipulation 的“perception”不能只围绕目标物体,支撑面本身也是操作状态的一部分。

第二类是触觉。Learning Tactile-Aware Quadrupedal Loco-Manipulation Policies 让高层扩散策略预测视觉—触觉目标,再由低层跟踪触觉命令;在 20 次 rollout 的接触任务中,加入低层触觉跟踪把插入成功率从 0.70 提到 0.85,完整重定向加插入从 0.60 提到 0.80。TAC-LOCO 更进一步,把触觉阵列、腿、臂和夹爪放入一个动态全身策略。在相同命令的 10 次硬件试验中,其成功率为 90%,固定夹爪 Deep WBC 基线为 50%。这意味着触觉正在从“操作策略的附加观测”变成“调节全身受力与抓持”的控制信号。

第三类是故障状态。FT-WBC 在未见过的锁死关节和极限放置高度下仍保持 70% 生存率,但任务成功只有 45%。这不是简单的性能下降,而是显式的 safety–completion trade-off:姿态模块限制激进俯仰,以防质心逃出退化后的支撑多边形。其边界是目前只处理单执行器故障,且任务多为静止操作或先移动后操作,尚非持续动态操作。

感知闭环仍是部署短板。ULTRA 在真实 G1 的稀疏目标跟随中,使用动捕物体状态的纵向/横向成功率为 80%/90%,换成机载深度点云后为 50%/60%。所以,控制器能否接受稀疏目标已经不是唯一问题;遮挡、点云提取和状态估计精度会直接决定全身策略是否可用。

四、长时序自主性:记忆和规划开始补课,但真实闭环证据仍弱

低层动作稳定后,机器人会遇到新的失败:物体离开视野后忘记位置、前一阶段的小误差累积到下一阶段、规划器给出不可执行姿态,或者论文只分别测了规划和控制。

SERF 用时空神经点图同时记录环境和机器人状态,在三个 BEHAVIOR-1K 仿真任务上把图像 VLA 的任务进度 40.7/43.0/48.4 提升到 63.5/60.1/52.5。它说明显式记忆对大空间、多物体任务有价值,但当前实验依赖 simulator instance label、先验地图和刚体跟踪,距离真实在线建图仍有明显落差。

模型式任务与运动规划也在向全身动力学靠拢。Task and Motion Planning for Humanoid Loco-manipulation 把 PDDL/MCTS 与全身动力学、接触和力矩约束结合,但双箱任务 200 次扩展平均仍需 52.3 秒,真实部署留作未来工作。另一边,Kitchen-R 虽有超过 500 条语言指令和约 2700 条轨迹,但为了隔离规划误差,执行评测始终使用 ground-truth plan;因此其模块得分不能直接解释为端到端自治成功率。

这里的研究空白很明确:需要同时公开语言任务成功、规划可行率、全身执行成功、恢复行为和累计时延,而不是只在其中一个层级给高分。

五、Sim-to-Real:出现正例,但“窄任务成功”仍不能外推为可靠部署

近一年对仿真迁移的态度更成熟:一部分工作展示零样本成功,另一部分则把失败明确量化。

SIMPLE 提供 60 个任务、千余物体和 50 个场景;在两个零样本任务上,pick-and-place 从仿真 9/10 到真实 8/10,handover 从 10/10 到 8/10。这是积极信号,但真实测试只覆盖两个短任务。WHOLE-MoMa 用 WBC 生成数据再做 offline RL,真实 cupboard 成功 17/25,优于生成器 4/25 和行为克隆 8/25;然而将 marker-based state 换成普通 pose tracker 后又降到 10/25,表明“策略迁移成功”常常仍以高精度状态估计为前提。

反例同样重要。MoDeSuite 中,两种图像策略在仿真窗帘任务成功,但真实世界 10 次都没有完成;检索策略仅 2 次接触窗帘,行为克隆甚至未到达。另有部署分析显示,提高 MuJoCo 摩擦冲量比可以消除滑移,却反映的是牵引参数调节而非控制器内在鲁棒性(A Framework for Deploying Learning-based Quadruped Loco-Manipulation)。

因此,今后的 sim-to-real 结果至少应回答:真实任务是否与仿真同分布;成功是否依赖动捕、marker 或已知几何;是否报告多次试验与失败分解;是否覆盖可变摩擦、柔顺、遮挡和重置误差。否则,单个演示视频的证据强度很有限。

可操作的系统设计框架

层级近一年较可信的选择仍需警惕
数据少量全身锚点 + 同形态静态操作 + 人类/robot-free 数据 + 仿真扩增跨形态数据不自动提供新接触动作
高层语言/视觉策略输出稀疏目标、潜在动作或可执行轨迹长时序记忆、主动感知与可行性约束不足
低层统一全身策略或“在线规划 + 预训练全身策略”动作接口若先天解耦,会限制能力上限
感知同时建模目标、地形、触觉、故障和机器人自身空间状态动捕/marker 依赖会掩盖真实部署难度
评测分阶段指标 + 端到端成功 + 失败恢复 + 多次真实试验模块高分不能替代闭环自治证据

研究空白与下一步

第一,需要“原生全身自治”而非只统一控制:把语言任务、场景记忆、主动观察、接触状态和全身可执行性放进一个可恢复闭环。第二,需要跨论文可比较的分层指标,明确 perception、planning、whole-body tracking、contact 与 recovery 各自的失败概率。第三,需要从刚体和受控实验室扩展到可变形物、柔顺机构、未知摩擦和多人环境。第四,需要安全从 reward 设计走向可验证约束,特别是负载、跌倒、关节故障与人机接触下的风险界限。

结论

近一年 loco-manipulation 已经越过“把导航和抓取接起来”的早期阶段。全身动作接口、统一控制、异构数据共训、地形与触觉反馈都给出了真实硬件上的可验证进展;其中,数据来源的分工和高低层职责的重新划分,可能比单纯增大 VLA 更具普适价值。但当前最强系统仍高度依赖任务结构、精确状态或受控场景。真正决定下一阶段的,不是机器人能否完成一个新 demo,而是它在目标变化、感知失真、接触突变和阶段性失败之后,能否继续完成任务。

References

Accepted evidence

去重论文引用

  1. A Framework for Deploying Learning-based Quadruped Loco-ManipulationYadong Liu, Jianwei Liu, He Liang, Dimitrios Kanoulas · 2025
  2. EgoHumanoid: Unlocking In-the-Wild Loco-Manipulation with Robot-Free Egocentric DemonstrationModi Shi, Shijia Peng, Jin Chen, Haoran Jiang, Tianyu Li, Di Huang, Ping Luo, Hongyang Li et al. · 2026
  3. Embracing Bulky Objects with Humanoid Robots: Whole-Body Manipulation with Reinforcement LearningChunxin Zheng, Kai Chen, Zhihai Bi, Yulin Li, Liang Pan, Jinni Zhou, Haoang Li, Jun Ma · 2025
  4. FT-WBC: Learning Fault-Tolerant Whole-Body Control for Legged Loco-ManipulationYudong Zhong, Pengfei Mai, Sikai Guo, Jiahang Cao, Zhihai Bi, Qiuyue Liu, Ziyan Feng, Jinni Zhou et al. · 2026
  5. Humanoid Manipulation Interface: Humanoid Whole-Body Manipulation from Robot-Free DemonstrationsRuiqian Nai, Boyuan Zheng, Junming Zhao, Haodong Zhu, Sicong Dai, Zunhao Chen, Yihang Hu, Yingdong Hu et al. · 2026
  6. HumanoidMimicGen: Data Generation for Loco-Manipulation via Whole-Body PlanningKevin Lin, Ajay Mandlekar, Caelan Reed Garrett, Nikita Chernyadev, Yu Fang, Runyu Ding, Yuqi Xie, Justin Tran et al. · 2026
  7. Learning Tactile-Aware Quadrupedal Loco-Manipulation PoliciesPokuang Zhou, Yuhao Zhou, Quan Khanh Luu, Seungho Han, Heng Zhang, Binghao Huang, Yunzhu Li, Arash Ajoudani et al. · 2026
  8. Learning Terrain-Aware Whole-Body Control for Perceptive Legged Loco-ManipulationSikai Guo, Yudong Zhong, Guoyang Zhao, Botao Dang, Zhihai Bi, Jun Ma · 2026
  9. Mind and Motion Aligned: A Joint Evaluation IsaacSim Benchmark for Task Planning and Low-Level Policies in Mobile ManipulationNikita Kachaev, Andrei Spiridonov, Andrey Gorodetsky, Kirill Muravyev, Nikita Oskolkov, Aditya Narendra, Vlad Shakhuro, Dmitry Makarov et al. · 2025
  10. MLM: Learning Multi-task Loco-Manipulation Whole-Body Control for Quadruped Robot with ArmXin Liu, Bida Ma, Chenkun Qi, Yan Ding, Nuo Xu, Zhaxizhuoma, Guorong Zhang, Pengan Chen et al. · 2025
  11. MoDeSuite: Robot Learning Task Suite for Benchmarking Mobile Manipulation with Deformable ObjectsYuying Zhang, Kevin Sebastian Luck, Francesco Verdoja, Ville Kyrki, Joni Pajarinen · 2025
  12. OpenHLM: An Empirical Recipe for Whole-Body Humanoid Loco-ManipulationYingdong Hu, Haodong Zhu, Boyuan Zheng, Yihang Hu, Tong Zhang, Zunhao Chen, Junming Zhao, Ruiqian Nai et al. · 2026
  13. Risk-Aware Reinforcement Learning for Mobile ManipulationMichael Groom, James Wilson, Nick Hawes, Lars Kunze · 2026
  14. SERF: Spatiotemporal Environment and Robot Feature Map for Long-Horizon Mobile ManipulationSunghwan Kim, Byeonghyun Pak, Kehan Long, Yulun Tian, Nikolay Atanasov · 2026
  15. SIMPLE: Simulation-Based Policy Learning and Evaluation for Humanoid Loco-manipulationSonglin Wei, Zhenhao Ni, Jie Liu, Zhenyu Zhao, Junjie Ye, Hongyi Jing, Junkai Xia, Xiawei Liu et al. · 2026
  16. Sumo: Dynamic and Generalizable Whole-Body Loco-ManipulationJohn Z. Zhang, Maks Sorokin, Jan Brüdigam, Brandon Hung, Stephen Phillips, Dmitry Yershov, Farzad Niroui, Tong Zhao et al. · 2026
  17. TAC-LOCO: Unified Whole-Body Control for Quadrupedal TACtile-Informed LOCO-ManipulationMuqun Hu, Yuhao Zhou, Kabir Ray Malik, Chi Lin, Won Suk Lee, Yu She, Yan Gu · 2026
  18. Task and Motion Planning for Humanoid Loco-manipulationMichal Ciebielski, Victor Dhédin, Majid Khadiv · 2025
  19. ULTRA: Unified Multimodal Control for Autonomous Humanoid Whole-Body Loco-ManipulationXialin He, Sirui Xu, Xinyao Li, Runpei Dong, Liuyu Bian, Yu-Xiong Wang, Liang-Yan Gui · 2026
  20. Whole-Body Mobile Manipulation using Offline Reinforcement Learning on Sub-optimal ControllersSnehal Jauhri, Vignesh Prasad, Georgia Chalvatzaki · 2026
  21. WholeBodyVLA: Towards Unified Latent VLA for Whole-Body Loco-Manipulation ControlHaoran Jiang, Jin Chen, Qingwen Bu, Li Chen, Modi Shi, Yanjie Zhang, Delong Li, Chuanzhe Suo et al. · 2025
完整证据附录(21 条正式证据事件)
  • Time range: 2025-07-19 至 2026-07-19
  • Events: 21
  • 每个事件一节,标题即锚点;trace-map 中的 event 链接跳转到这里。

EA-LOCOMANIP-2026-0005

  • Claim: In simulation, the NSDF-conditioned policy maintained 100% success for standard-mass cylinders, cuboids and spheres, while removing NSDF produced 0% across those shapes.
  • Stance: support | Confidence: direct
  • Paper: 2509.13534 Embracing Bulky Objects with Humanoid Robots: Whole-Body Manipulation with Reinforcement Learning
  • Locator: IV-B Adaptability to Different Object Properties
  • Evidence: The shape sweep directly contrasts the full method with the no-NSDF ablation.
  • Quote: “First, when the NSDF module is incorporated, the policy achieves and maintains a 100% success rate across multiple object types, including cylinders, cuboids, and spheres, under standard mass (3 kg) and size conditions. This demonstrates strong sim-to-sim transfer performance. In contrast, the policy exhibits complete failure (0% success) across all shape categories without NSDF, highlighting the essential role of this module in capturing geometric and semantic attributes for stable manipulation”
  • Authors: chunxin-zheng; kai-chen; zhihai-bi; et al.

EA-LOCOMANIP-2026-0006

  • Claim: Removing the unified latent action model reduced success by 38.7 percentage points, indicating that action-free human video contributed useful priors in the evaluated tasks.
  • Stance: support | Confidence: direct
  • Paper: 2512.11047 WholeBodyVLA: Towards Unified Latent VLA for Whole-Body Loco-Manipulation Control
  • Locator: 4.3 How does action-free videos contribute to loco–manipulation?
  • Evidence: The ablation directly compares the full model with removal of unified latent learning.
  • Quote: “As shown in Table 4.2 , the full model improves success rate by 38.7%, indicating that unified latent learning extracts useful priors from action-free human videos and enhances downstream policy learning.”
  • Authors: haoran-jiang; jin-chen; qingwen-bu; et al.

EA-LOCOMANIP-2026-0008

  • Claim: In a 15-minute comparison, HuMI collected 62 episodes versus 28 for TWIST2, with 96.7% versus 64.3% acceptance; time per acceptable episode fell to 30.0% of TWIST2.
  • Stance: support | Confidence: direct
  • Paper: 2602.06643 Humanoid Manipulation Interface: Humanoid Whole-Body Manipulation from Robot-Free Demonstrations
  • Locator: VI Data Collection Efficiency
  • Evidence: The timed comparison jointly reports quantity, quality and accepted-data cost.
  • Quote: “HuMI yields substantially higher throughput, collecting 62 episodes versus 28 with TWIST2, while also achieving a higher acceptance rate ( 96.7% vs. 64.3% ). HuMI’s streamlined, robot-free workflow further reduces the average time per acceptable episode to 30.0% of that of TWIST2”
  • Authors: ruiqian-nai; boyuan-zheng; junming-zhao; et al.

EA-LOCOMANIP-2026-0009

  • Claim: With 100 robot and 300 human demonstrations, co-training scored 78% versus 59% for robot-only in-domain, and 82% versus 31% under generalization.
  • Stance: support | Confidence: direct
  • Paper: 2602.10106 EgoHumanoid: Unlocking In-the-Wild Loco-Manipulation with Robot-Free Egocentric Demonstration
  • Locator: IV-B Will human data improve humanoid loco-manipulation?
  • Evidence: The main comparison directly reports both in-domain and generalization gaps.
  • Quote: “For in-domain evaluation, co-training achieves 78% average score versus 59% for robot-only. The gap widens remarkably in generalization settings: co-training reaches 82% compared to 31% for robot-only.”
  • Authors: modi-shi; shijia-peng; jin-chen; et al.

EA-LOCOMANIP-2026-0019

  • Claim: On hardware, Sumo uprighted a 15 kg tire—heavier than Spot arm's stated 11 kg lifting capacity—in 10/10 trials, averaging 9.2±4.7 seconds.
  • Stance: support | Confidence: direct
  • Paper: 2604.08508 Sumo: Dynamic and Generalizable Whole-Body Loco-Manipulation
  • Locator: page 8
  • Evidence: The hardware case study reports object mass, nominal arm capacity, success and completion time.
  • Quote: “Tire Upright:In Fig. 7 (a), the robot is tasked to upright a tire of15kg, which is heavier than the maximum lifting capacity of11kg of the Spot arm. Additionally, the rubber tire presents hard-to-model geometries and friction properties, causing a large sim-to-real gap. Using a combination of its arm, torso, and legs, Sumo enables the robot to complete the task10out of10trials with an average completion time of9.2±4.7s, Tab. II.”
  • Authors: john-z-zhang; maks-sorokin; jan-brdigam; et al.

EA-LOCOMANIP-2026-0011

  • Claim: On real cupboard opening, WHOLE-MoMa succeeded in 17/25 trials (68%), versus 4/25 for its WBC generator and 8/25 for behavior cloning.
  • Stance: support | Confidence: direct
  • Paper: 2604.12509 Whole-Body Mobile Manipulation using Offline Reinforcement Learning on Sub-optimal Controllers
  • Locator: VI-B Real-World Results
  • Evidence: The real-world table separates full success across the three methods.
  • Quote: “WHOLE-MoMa achieves 17/25 (68%), demonstrating substantial sim-to-real transfer, with the highest grasping success (22/25) and articulation success (17/22 given grasping).”
  • Authors: snehal-jauhri; vignesh-prasad; georgia-chalvatzaki

EA-LOCOMANIP-2026-0012

  • Claim: Adding tactile-command tracking at the low level raised insertion success from 0.70 to 0.85, full reorientation-plus-insertion from 0.60 to 0.80, and valve tightening from 0.80 to 0.85.
  • Stance: support | Confidence: direct
  • Paper: 2604.27224 Learning Tactile-Aware Quadrupedal Loco-Manipulation Policies
  • Locator: IV-B Experimental Results and Analyze
  • Evidence: The paper compares variants with the same tactile-aware high level but different low-level tactile tracking.
  • Quote: “Finally, comparing Baseline 3 (P3) with our full method demonstrates the benefit of incorporating tactile commands into the low-level policy: our method further increases the success rate from 0.70 to 0.85 on Task 1 insertion (+0.15), from 0.60 to 0.80 on Task 1 whole (+0.20), and from 0.80 to 0.85 on Task 2 (+0.05), and achieves 1.00 on Task 3 (compared to 0.20 for P1).”
  • Authors: pokuang-zhou; yuhao-zhou; quan-khanh-luu; et al.

EA-LOCOMANIP-2026-0013

  • Claim: Across nine simulated tasks, data generated from one source demonstration raised average policy performance from 0.33 for DexMimicGen+ to 0.89 for HumanoidMimicGen.
  • Stance: support | Confidence: direct
  • Paper: 2605.27724 HumanoidMimicGen: Data Generation for Loco-Manipulation via Whole-Body Planning
  • Locator: 6.2 HumanoidMimicGen Capabilities
  • Evidence: The main comparison reports downstream policy performance averaged over all nine tasks.
  • Quote: “From Table 1 , HumanoidMimicGen improves policy performance over all baselines in all tasks. Averaged over nine tasks, data generated by HumanoidMimicGen from a single source human demonstration increases policy performance from 0.33 (DexMimicGen+) to 0.89”
  • Authors: kevin-lin; ajay-mandlekar; caelan-reed-garrett; et al.

EA-LOCOMANIP-2026-0014

  • Claim: In a staged real long-horizon task, TA-WBC completed five consecutive bottle-pick, stair-climb, disposal and return runs without falls or stumbles.
  • Stance: support | Confidence: direct
  • Paper: 2605.31343 Learning Terrain-Aware Whole-Body Control for Perceptive Legged Loco-Manipulation
  • Locator: IV-C 2 Long-Horizon Loco-Manipulation
  • Evidence: The real-world section explicitly reports repeated completion of the full multi-stage route.
  • Quote: “The result shows that the legged manipulator with TA-WBC can successfully finish this long-horizon challenging task 5 times in a row without any falls or stumbles, verifying the robustness of our framework.”
  • Authors: sikai-guo; yudong-zhong; guoyang-zhao; et al.

EA-LOCOMANIP-2026-0020

  • Claim: On four tasks held out from whole-body teleoperation, stationary same-embodiment co-training raised average task progress from 33% to 87%, close to a 94% 12-task teleoperation oracle.
  • Stance: support | Confidence: direct
  • Paper: 2606.22174 OpenHLM: An Empirical Recipe for Whole-Body Humanoid Loco-Manipulation
  • Locator: 1 Introduction
  • Evidence: The paper overview reports the held-out-task gain and the full-teleoperation oracle.
  • Quote: “Quantitatively, OpenHLM reaches 89% average task progress on the 8 training tasks of our HLM-12 benchmark; on the remaining 4 held-out tasks that whole-body teleop never covers, co-training with cheaper data lifts task progress from 33% to 87%, closing most of the gap to a 12-task oracle (94%).”
  • Authors: yingdong-hu; haodong-zhu; boyuan-zheng; et al.

EA-LOCOMANIP-2026-0021

  • Claim: In 10 matched hardware trials, tactile-informed TAC-LOCO achieved 90% dynamic loco-manipulation success versus 50% for Deep WBC with a fixed gripper.
  • Stance: support | Confidence: direct
  • Paper: 2607.10132 TAC-LOCO: Unified Whole-Body Control for Quadrupedal TACtile-Informed LOCO-Manipulation
  • Locator: 6.5 Baseline comparison
  • Evidence: The hardware baseline comparison isolates learned grasp regulation under the same command set.
  • Quote: “We conduct 10 hardware trials using the same set of loco-manipulation commands as in Sec. 6 . Table 4 compares the success rates of our policy and the baseline, showing that our tactile-informed policy achieves a substantially higher success rate. Figure 11 shows that the baseline suffers from gradual object slip during the task. Since the gripper width remains fixed, the policy cannot actively suppress slip once the external force changes. Table 4: Deployment success rate comparison with baseli”
  • Authors: muqun-hu; yuhao-zhou; kabir-ray-malik; et al.

EA-LOCOMANIP-2026-0002

  • Claim: On two real-world tasks, the same controller achieved 98% and 100% success under teleoperation, versus 80% and 85% when driven by a diffusion policy.
  • Stance: conditional | Confidence: direct
  • Paper: 2508.10538 MLM: Learning Multi-task Loco-Manipulation Whole-Body Control for Quadruped Robot with Arm
  • Locator: IV-C Real-world Experiments
  • Evidence: Table V directly separates low-level tracking results under teleoperation and diffusion-policy trajectory generation.
  • Quote: “TABLE V: Real-world experiment results. (cm / radian / %) EE pos err EE rot err Success Rate Teleoperation unplug charger 1.39 0.054 98 open container 1.02 0.047 100 DP unplug charger 1.37 0.039 80 open container 1.24 0.036 85”
  • Authors: xin-liu; bida-ma; chenkun-qi; et al.

EA-LOCOMANIP-2026-0003

  • Claim: For two-box pick-and-place, the solver found the first goal-satisfying feasible plan after 30 of 200 tree expansions, with an average solve time of 52.3 seconds.
  • Stance: conditional | Confidence: direct
  • Paper: 2508.14099 Task and Motion Planning for Humanoid Loco-manipulation
  • Locator: IV-C Example Solutions
  • Evidence: The example-solution section reports both search effort and computation for the modeled task.
  • Quote: “The two-box pick and place tree search was expanded 200 times with an average solve time of 52.3 seconds. Note that using an NLP solver that exploits the sparsity in time of trajectory optimization, such as the SQP solver in Acados [ 20 ] , could yield a significant speedup w.r.t. IPOPT. A variety of solutions were generated, with the first feasible solution satisfying the goal being found in 30 iterations.”
  • Authors: michal-ciebielski; victor-dhdin; majid-khadiv

EA-LOCOMANIP-2026-0018

  • Claim: On real G1 sparse-goal following, MoCap object state achieved 80% vertical and 90% lateral success, while egocentric depth achieved 50% and 60%, respectively.
  • Stance: conditional | Confidence: direct
  • Paper: 2603.03279 ULTRA: Unified Multimodal Control for Autonomous Humanoid Whole-Body Loco-Manipulation
  • Locator: V-E Real-World Deployment
  • Evidence: The real-world table separates external-state and onboard egocentric control modes.
  • Quote: “TABLE IV : Real-world success rates on the OMOMO subset using a Unitree G1 humanoid. Each task is evaluated over two trials. MoCap provides object pose tracking for non-egocentric control modes, while the egocentric setting relies only on onboard sensing. MoCap is used for success evaluation in all settings. Dense reference tracking is direction-agnostic and thus reported as a single success rate. Setting Vertical Lateral Dense Reference Tracking 73% (19/26) Sparse Goal Following (MoCap) 80% (8/”
  • Authors: xialin-he; sirui-xu; xinyao-li; et al.

EA-LOCOMANIP-2026-0010

  • Claim: The study reports a depth-only mobile-manipulation policy whose risk sensitivity can be adjusted at runtime while retaining task performance comparable to risk-neutral methods in simulation.
  • Stance: conditional | Confidence: direct
  • Paper: 2603.04579 Risk-Aware Reinforcement Learning for Mobile Manipulation
  • Locator: VI Conclusion
  • Evidence: The conclusion summarizes risk-aware student competence and transfer through imitation learning.
  • Quote: “Our approach trains risk-aware visuomotor policies for mobile manipulation conditioned on egocentric depth observations with runtime-adjustable risk sensitivity, achieving comparable performance to risk-neutral methods, and we show learnt risk-aware behaviours can be successfully transferred through Imitation Learning.”
  • Authors: michael-groom; james-wilson; nick-hawes; et al.

EA-LOCOMANIP-2026-0015

  • Claim: In zero-shot transfer, pick-and-place scored 9/10 in simulation and 8/10 on hardware; handover scored 10/10 in simulation and 8/10 on hardware.
  • Stance: conditional | Confidence: direct
  • Paper: 2606.08278 SIMPLE: Simulation-Based Policy Learning and Evaluation for Humanoid Loco-manipulation
  • Locator: 4 Experiments
  • Evidence: The transfer table reports paired simulation and real success counts; in the recovered HTML it is attached to the parent experiments section.
  • Quote: “Task Sim Eval Real Eval Pick & Place 9/10 = 0.90 8/10 = 0.80 Handover 10/10 = 1.00 8/10 = 0.80 Table 5: Zero-Shot Sim-to-Real Transfer. Success rates of a single policy fine-tuned exclusively on simulation data, evaluated both in the simulator and directly in the real world.”
  • Authors: songlin-wei; zhenhao-ni; jie-liu; et al.

EA-LOCOMANIP-2026-0016

  • Claim: Across three simulated BEHAVIOR-1K tasks, SERF achieved mean task progress of 63.5, 60.1 and 52.5, versus 40.7, 43.0 and 48.4 for the fine-tuned image-only PI0.5 baseline.
  • Stance: conditional | Confidence: direct
  • Paper: 2606.12956 SERF: Spatiotemporal Environment and Robot Feature Map for Long-Horizon Mobile Manipulation
  • Locator: 5 Long-Horizon Mobile Manipulation
  • Evidence: The main table reports mean task progress for the full system and image-only baseline across three tasks.
  • Quote: “The full SERF policy achieves the highest task progress on all three evaluated tasks.”
  • Authors: sunghwan-kim; byeonghyun-pak; kehan-long; et al.

EA-LOCOMANIP-2026-0017

  • Claim: Under an unseen locked-joint fault at the most demanding placement height, FT-WBC retained 70% survival but only 45% task success, reflecting an explicit survival-first posture policy.
  • Stance: conditional | Confidence: direct
  • Paper: 2606.24466 FT-WBC: Learning Fault-Tolerant Whole-Body Control for Legged Loco-Manipulation
  • Locator: 4.3 Real-World Deployment
  • Evidence: The real-world analysis directly attributes reduced completion to conservative posture limits that preserve the degraded support polygon.
  • Quote: “At the extreme height, our framework still guarantees a 70% survival rate even under severe locked conditions, demonstrating strong adaptation to out-of-distribution actuator failures. While the task success rate drops to 45%, this occurs because the PAM conservatively restricts the extreme upward pitch to prevent the CoM from escaping the degraded support polygon.”
  • Authors: yudong-zhong; pengfei-mai; sikai-guo; et al.

EA-LOCOMANIP-2026-0001

  • Claim: On the real curtain task, neither image policy completed any of 10 trials; retrieval reached the curtain in 2/10, while behavior cloning did not reach it.
  • Stance: limit | Confidence: direct
  • Paper: 2507.21796 MoDeSuite: Robot Learning Task Suite for Benchmarking Mobile Manipulation with Deformable Objects
  • Locator: IV-C Deployment on Real Robot
  • Evidence: The deployment paragraph contrasts simulation success with physical failure.
  • Quote: “While the policies are successful in simulation, neither is able to complete the task in the real world. Specifically, the retrieval-based method managed to approach and make contact with the curtain in 2 out of 10 trials, whereas the behavior cloning (BC) policy failed to even reach the curtain.”
  • Authors: yuying-zhang; kevin-sebastian-luck; francesco-verdoja; et al.

EA-LOCOMANIP-2026-0004

  • Claim: Kitchen-R's reported execution evaluation always uses the ground-truth plan to isolate execution from planning error, so its module results are not direct evidence of end-to-end autonomy.
  • Stance: limit | Confidence: direct
  • Paper: 2508.15663 Mind and Motion Aligned: A Joint Evaluation IsaacSim Benchmark for Task Planning and Low-Level Policies in Mobile Manipulation
  • Locator: IV METRICS
  • Evidence: The metrics section explicitly states that execution does not use the predicted plan.
  • Quote: “to isolate planning errors from execution errors, the ground-truth plan is always used for execution.”
  • Authors: nikita-kachaev; andrei-spiridonov; andrey-gorodetsky; et al.

EA-LOCOMANIP-2026-0007

  • Claim: Cross-simulator smoothness was not a reliable robustness signal: MuJoCo drifted under default friction, while near-zero stop error under tuned friction came from unrealistically high tangential impedance.
  • Stance: limit | Confidence: direct
  • Paper: 2512.18938 A Framework for Deploying Learning-based Quadruped Loco-Manipulation
  • Locator: 4.1.3 MuJoCo (tuned friction)
  • Evidence: The cross-simulator study attributes apparently improved MuJoCo results to contact-parameter tuning rather than policy quality.
  • Quote: “When evaluated with artificially enlarged impulse ratio ( impratio=100 ), sliding was nearly eliminated and the neutral/stop phases reached near-zero error ( Table ˜ 2 ). This confirms that apparent smoothness reflects traction tuning rather than intrinsic robustness.”
  • Authors: yadong-liu; jianwei-liu; he-liang; et al.

References

Interactive Wiki

进入交互阅读

静态页采用默认版本;v0、v0、v0可在同一交互 Wiki 中阅读,不建立相互竞争的索引页。

Method and limits

研究方法、证据边界与引用说明

本页由当前结算的证据层自动派生,不改写已结算 run。论文只有在完整非 OCR 全文、结构化论文笔记和 claim-support audit 通过后,才进入接受证据。

页面结论是对所列论文与研究窗口的综合,不等同于无限外推。引用本专题时,请同时保留本页 canonical 地址、访问日期,并在需要时回溯原论文。