https://rubatotree.github.io/blog/tags/%E8%AE%BA%E6%96%87%E9%80%9F%E8%A7%88/Recent content in 论文速览 on Umbrella Coffeehttps://rubatotree.github.io/blog/images/og-default.pnghttps://rubatotree.github.io/blog/images/og-default.pngHugoen-usTue, 04 Aug 2026 00:00:00 +0800https://rubatotree.github.io/blog/posts/icra2025-data-expansion-paper-notes/Tue, 04 Aug 2026 00:00:00 +0800https://rubatotree.github.io/blog/posts/icra2025-data-expansion-paper-notes/<p>本文联合整理 ICRA 2025、IROS 2025 与 ICRA 2026 的数据扩展工作。ICRA 2025 Proceedings 收录 1,604 篇论文并补全 1,602 篇摘要，自动召回 279 篇候选并初审 32 篇；IROS 2025 Proceedings 收录 1,985 篇论文并补全 1,983 篇摘要，从 309 篇候选初审 31 篇。ICRA 2026 则依据官方 PaperCept 程序的 <strong>ICRA 2026 peer-reviewed conference presentations</strong>：2,951 条程序行中排除 131 条非同行评议、非出版物的 Late Breaking Results，保留 2,820 条展示（2,604 个 interactive、216 个 oral）；其中 2,639 条提供官方关键词与摘要。复审把官方摘要纳入生命周期匹配、将 5 分项纳入人工复核带，当前快照自动召回 451 篇候选并人工审计 35 篇。随后以同一“数据–下游 Policy/控制”门槛复核渲染与光照方向，额外补入 ICRA 2025 的 2 篇与 IROS 2025 的 3 篇；正文当前分别覆盖 34、34、35 篇。三部分合计审计 6,409 条正式论文或同行评议展示，在同一主题框架下比较机器人数据如何被采集、转换、合成、筛选和复用。</p>https://rubatotree.github.io/blog/posts/corl2025-oral-paper-notes/Mon, 03 Aug 2026 00:00:00 +0800https://rubatotree.github.io/blog/posts/corl2025-oral-paper-notes/<p>本文使用 AI 工具整理 <a href="https://openreview.net/group?id=robot-learning.org/CoRL/2025/Conference#tab-accept-oral" target="_blank" rel="noopener noreferrer"><span style="color: #59a4ff;"><span style="text-decoration: underline">CoRL 2025 Oral</span></span></a> 的 42 篇论文。为了避免把列表写成摘要搬运，下面按研究问题重新分为六组，并分别概括方法、主要结果与值得留意的边界。</p> <p>需要说明的是，这是一份基于 OpenReview 标题与摘要的“地图式速览”，适合用来筛选待读论文，不等价于逐篇精读或实验复现。文中的“提升”“达到”等结果均指作者报告的结果。</p> <h2 id="loc-1">总览：Oral 在解决什么问题</h2> <p>这 42 篇论文呈现出一条很清楚的主线：机器人学习的瓶颈正在从“能否学会一个任务”转向“能否低成本地扩展到新任务、新环境和新本体”。对应地，工作重点也从单纯设计更大的 Policy，转向数据生产、跨本体迁移、推理效率、真实世界评测与失败恢复等完整系统问题。</p> <p>其中有四个趋势尤其明显。第一，真实机器人示教不再是唯一的数据来源：人类视频、手机扫描、生成模型、仿真与低成本外骨骼都在被转化为训练数据。第二，Diffusion Policy 与 VLA 已成为基础组件，研究开始优化其执行速度、在线适应与推理机制。第三，触觉、音频、力与 LiDAR 等模态被纳入闭环控制，而不只是作为附加观测。第四，研究者开始正面处理评测偏差、分布外失败和开放世界泛化，说明领域正在从 Demo 走向可部署系统。</p> <h2 id="loc-2">1. 通用策略、推理、评测与安全</h2> <p>这一组关注的不是某项具体技能，而是通用机器人策略如何获得开放世界泛化能力，以及我们如何可信地评价和保护这些策略。</p> <h3 id="loc-3">1. <span role="math"><svg class="typst-frame" style="overflow: visible; width: 1.6599em; height: 0.683em;" viewBox="0 0 26.89038 11.0646" width="26.89038pt" height="11.0646pt" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" xmlns:h5="http://www.w3.org/1999/xhtml"><g><g class="typst-text" transform="matrix(1 0 0 -1 0 11.0646)"><use xlink:href="#g26BF746A2A3102A61C8A8433F06F3F20" x="0" y="0" fill="#000000" fill-rule="nonzero"/></g><g class="typst-text" transform="matrix(1 0 0 -1 9.234 15.066)"><use xlink:href="#gF479BA07BDB4B1D1F2BD759B5428359D" x="0" y="0" fill="#000000" fill-rule="nonzero"/></g><g class="typst-text" transform="matrix(1 0 0 -1 15.68646 15.066)"><use xlink:href="#gDCC78161D122F6A91A5B72AC0D240F83" x="0" y="0" fill="#000000" fill-rule="nonzero"/></g><g class="typst-text" transform="matrix(1 0 0 -1 19.530720000000002 15.066)"><use xlink:href="#gC4BE5334B99364CC275FE24AD40A6813" x="0" y="0" fill="#000000" fill-rule="nonzero"/></g></g><defs id="glyph"><symbol id="g26BF746A2A3102A61C8A8433F06F3F20" overflow="visible"><path d="M 0 0m 9.1854 6.5934 c 0 0.38880014 -0.34019947 0.38880014 -0.64799976 0.38880014 h -5.427 c -0.3564 0 -0.9720001 0 -1.6848001 -0.76139975 c -0.56700003 -0.6318002 -0.98820007 -1.3770003 -0.98820007 -1.4580002 c 0 0 0 -0.16200018 0.19439998 -0.16200018 c 0.12960005 0 0.162 0.06480026 0.25920004 0.19440031 c 0.7938 1.2473998 1.7334001 1.2473998 2.0574 1.2473998 h 0.92340016 c -0.5184002 -1.9601998 -1.3932002 -3.9204001 -2.0736003 -5.3946 c -0.12959993 -0.243 -0.12959993 -0.27539998 -0.12959993 -0.3888 c 0 -0.3078 0.25919998 -0.4374 0.4698 -0.4374 c 0.48600006 0 0.6156001 0.45360002 0.80999994 1.053 c 0.22679996 0.7452 0.22679996 0.7776 0.4374001 1.5876 l 0.9072001 3.5802002 h 1.8305998 c -0.5345998 -2.3976002 -0.6803999 -3.0942001 -0.6803999 -4.1796002 c 0 -0.24300003 0 -0.6804 0.12960005 -1.2312 c 0.16200018 -0.7128 0.34019995 -0.81 0.5832 -0.81 c 0.32399988 0 0.6641998 0.29160002 0.6641998 0.6156 c 0 0.097200006 0 0.12960002 -0.09719992 0.35639998 c -0.4698 1.1664001 -0.4698 2.2194 -0.4698 2.6729999 c 0 0.8585999 0.11339998 1.7334001 0.29160023 2.5758002 h 1.8468003 c 0.2105999 0 0.7937994 0 0.7937994 0.55079985 Z "/></symbol><symbol id="gF479BA07BDB4B1D1F2BD759B5428359D" overflow="visible"><path d="M 0 0m 5.85144 3.61746 c 0 1.2474 -0.14741993 2.1432598 -0.66906023 2.9370596 c -0.3515396 0.5216403 -1.0546198 0.97524023 -1.9618199 0.97524023 c -2.6308799 0 -2.6308799 -3.09582 -2.6308799 -3.9122999 c 0 -0.81648016 0 -3.84426 2.6308799 -3.84426 c 2.63088 0 2.63088 3.0277798 2.63088 3.84426 Z m -1.03194 0.14741993 c 0 -0.83915997 0 -1.6783199 -0.14741993 -2.39274 c -0.22679996 -1.03194 -0.99792004 -1.28142 -1.4515202 -1.28142 c -0.5216398 0 -1.2133799 0.30618 -1.4401798 1.2360599 c -0.15876007 0.66906 -0.15876007 1.59894 -0.15876007 2.4380999 c 0 0.82782006 0 1.6896598 0.17009997 2.31336 c 0.2381401 0.8958597 0.9639001 1.1339998 1.4288399 1.1339998 c 0.61236 0 1.20204 -0.3742199 1.4061601 -1.03194 c 0.18143988 -0.61236 0.19278002 -1.4288397 0.19278002 -2.4154198 Z "/></symbol><symbol id="gDCC78161D122F6A91A5B72AC0D240F83" overflow="visible"><path d="M 0 0m 2.57418 0.64638 c 0 0.40824002 -0.34019995 0.6577199 -0.64637995 0.6577199 c -0.37422 0 -0.65771997 -0.30617994 -0.65771997 -0.64637995 c 0 -0.40823996 0.34019995 -0.65771997 0.64637995 -0.65771997 c 0.37422 0 0.65771997 0.30618 0.65771997 0.64638 Z "/></symbol><symbol id="gC4BE5334B99364CC275FE24AD40A6813" overflow="visible"><path d="M 0 0m 5.7267 2.27934 c 0 1.28142 -0.97523975 2.46078 -2.3813999 2.46078 c -0.49896002 0 -1.08864 -0.124740124 -1.5876 -0.55566025 v 2.14326 c 0.57834005 -0.14741993 0.91854 -0.14741993 1.0999799 -0.14741993 c 1.49688 0 2.3814 1.0205998 2.3814 1.1907 c 0 0.124740124 -0.079380035 0.15876007 -0.13608027 0.15876007 c 0 0 -0.04535961 0 -0.0907197 -0.034019947 c -0.27216005 -0.10206032 -0.8731799 -0.32886028 -1.7123401 -0.32886028 c -0.3175199 0 -0.91854 0.022680283 -1.6556399 0.30618 c -0.11339998 0.05670023 -0.14742005 0.05670023 -0.14742005 0.05670023 c -0.14741993 0 -0.14741993 -0.124740124 -0.14741993 -0.30618 v -3.3453 c 0 -0.19278002 0 -0.32886004 0.18144 -0.32886004 c 0.10205996 0 0.11339998 0.022680044 0.22679996 0.15876007 c 0.48761988 0.62370014 1.1793599 0.71441984 1.5762599 0.71441984 c 0.68040013 0 0.98658013 -0.54431987 1.0432799 -0.6350398 c 0.20412016 -0.37422013 0.27216005 -0.80514 0.27216005 -1.4628601 c 0 -0.34019995 0 -1.0092599 -0.34019995 -1.50822 c -0.28349972 -0.40824 -0.77111983 -0.6804 -1.3267798 -0.6804 c -0.7371001 0 -1.49688 0.40824 -1.78038 1.15668 c 0.43092 -0.034020066 0.64637995 0.24947989 0.64637995 0.5556599 c 0 0.48762012 -0.41957998 0.57834005 -0.5669999 0.57834005 c 0 0 -0.56700003 0 -0.56700003 -0.61236 c 0 -1.0206 0.92988 -2.0412 2.29068 -2.0412 c 1.4515197 0 2.7215998 1.0773 2.7215998 2.50614 Z "/></symbol></defs></svg></span>: a Vision-Language-Action Model with Open-World Generalization [<a href="https://openreview.net/forum?id=vlhoswksBO" target="_blank" rel="noopener noreferrer"><span style="color: #59a4ff;"><span style="text-decoration: underline">OpenReview</span></span></a>]</h3> <p><span role="math"><svg class="typst-frame" style="overflow: visible; width: 1.6599em; height: 0.683em;" viewBox="0 0 22.408649999999998 9.2205" width="22.408649999999998pt" height="9.2205pt" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" xmlns:h5="http://www.w3.org/1999/xhtml"><g><g class="typst-text" transform="matrix(1 0 0 -1 0 9.2205)"><use xlink:href="#g21E0DFDE6212E4409ACBE5D7BF48012D" x="0" y="0" fill="#000000" fill-rule="nonzero"/></g><g class="typst-text" transform="matrix(1 0 0 -1 7.695 12.555000000000001)"><use xlink:href="#g53818CE14381D89ADAFC63CFBFE9057F" x="0" y="0" fill="#000000" fill-rule="nonzero"/></g><g class="typst-text" transform="matrix(1 0 0 -1 13.072049999999999 12.555000000000001)"><use xlink:href="#g9997D3520E6EC99978E01AC34B3B1991" x="0" y="0" fill="#000000" fill-rule="nonzero"/></g><g class="typst-text" transform="matrix(1 0 0 -1 16.275599999999997 12.555000000000001)"><use xlink:href="#gBCBEDF1D013990426CBDCE52923425D6" x="0" y="0" fill="#000000" fill-rule="nonzero"/></g></g><defs id="glyph"><symbol id="g21E0DFDE6212E4409ACBE5D7BF48012D" overflow="visible"><path d="M 0 0m 7.6545 5.4945 c 0 0.32399988 -0.28349972 0.32399988 -0.53999996 0.32399988 h -4.5225 c -0.29699993 0 -0.80999994 0 -1.4039999 -0.6345 c -0.4725001 -0.52649975 -0.82350004 -1.1475 -0.82350004 -1.2149999 c 0 0 0 -0.13499999 0.16199997 -0.13499999 c 0.10800004 0 0.13500005 0.0539999 0.21600002 0.16199994 c 0.66150004 1.0395 1.4445 1.0395 1.7145 1.0395 h 0.7695 c -0.43199992 -1.6334999 -1.161 -3.267 -1.728 -4.4955 c -0.10799992 -0.20250002 -0.10799992 -0.22950003 -0.10799992 -0.324 c 0 -0.2565 0.21599996 -0.36450002 0.3915 -0.36450002 c 0.40499997 0 0.513 0.37800002 0.67499995 0.87750006 c 0.18900013 0.621 0.18900013 0.64800006 0.36450005 1.323 l 0.75600004 2.9835 h 1.5255 c -0.4454999 -1.9979999 -0.5669999 -2.5785 -0.5669999 -3.483 c 0 -0.20249999 0 -0.567 0.1079998 -1.026 c 0.13500023 -0.594 0.2835002 -0.675 0.48600006 -0.675 c 0.26999998 0 0.5535002 0.24300002 0.5535002 0.513 c 0 0.081 0 0.10800001 -0.08100033 0.29700002 c -0.3915 0.97200006 -0.3915 1.8495002 -0.3915 2.2275002 c 0 0.7154999 0.094500065 1.4444997 0.24300003 2.1464999 h 1.539 c 0.1755004 0 0.6615 0 0.6615 0.4590001 Z "/></symbol><symbol id="g53818CE14381D89ADAFC63CFBFE9057F" overflow="visible"><path d="M 0 0m 4.8761997 3.01455 c 0 1.0395 -0.12284994 1.7860501 -0.55754995 2.44755 c -0.29294968 0.4347 -0.87884974 0.8126998 -1.6348498 0.8126998 c -2.1924 0 -2.1924 -2.57985 -2.1924 -3.2602499 c 0 -0.68040013 0 -3.2035499 2.1924 -3.2035499 c 2.1923997 0 2.1923997 2.5231497 2.1923997 3.2035499 Z m -0.8599496 0.12284994 c 0 -0.69930005 0 -1.3986 -0.12285018 -1.9939499 c -0.18900013 -0.85995007 -0.83159995 -1.06785 -1.2096 -1.06785 c -0.4347 0 -1.01115 0.25515 -1.20015 1.0300499 c -0.13230002 0.55754995 -0.13230002 1.3324499 -0.13230002 2.03175 c 0 0.6898501 0 1.4080498 0.14174998 1.9277999 c 0.19845009 0.7465501 0.8032501 0.9450002 1.1907 0.9450002 c 0.5102999 0 1.0016999 -0.31185007 1.1717999 -0.85995007 c 0.15119982 -0.51030016 0.16065025 -1.1907 0.16065025 -2.01285 Z "/></symbol><symbol id="g9997D3520E6EC99978E01AC34B3B1991" overflow="visible"><path d="M 0 0m 2.14515 0.53865 c 0 0.3402 -0.28349996 0.54810005 -0.53865004 0.54810005 c -0.31184995 0 -0.5481 -0.25515008 -0.5481 -0.53865004 c 0 -0.3402 0.28350008 -0.5481 0.53865004 -0.5481 c 0.31184995 0 0.5481 0.25515 0.5481 0.53865 Z "/></symbol><symbol id="gBCBEDF1D013990426CBDCE52923425D6" overflow="visible"><path d="M 0 0m 4.7722497 1.89945 c 0 1.06785 -0.8126998 2.0506501 -1.9844997 2.0506501 c -0.4158001 0 -0.9072001 -0.10395002 -1.3230001 -0.46305013 v 1.7860501 c 0.48195004 -0.12284994 0.7654501 -0.12284994 0.91664994 -0.12284994 c 1.2474 0 1.9845002 0.8505001 1.9845002 0.99224997 c 0 0.10395002 -0.06615019 0.1322999 -0.11339998 0.1322999 c 0 0 -0.037800312 0 -0.07560015 -0.028349876 c -0.22679996 -0.085050106 -0.7276499 -0.27405024 -1.42695 -0.27405024 c -0.26460004 0 -0.76545 0.018899918 -1.3797 0.25515032 c -0.094500065 0.047249794 -0.12285006 0.047249794 -0.12285006 0.047249794 c -0.12284994 0 -0.12284994 -0.10395002 -0.12284994 -0.25514984 v -2.78775 c 0 -0.16065001 0 -0.27405 0.15119994 -0.27405 c 0.085050106 0 0.094500065 0.018899918 0.18900001 0.1322999 c 0.40635002 0.5197501 0.9828001 0.59535 1.3135501 0.59535 c 0.5669999 0 0.82214975 -0.45359993 0.8693998 -0.5292001 c 0.17009997 -0.31184983 0.22679996 -0.67094994 0.22679996 -1.2190499 c 0 -0.28349996 0 -0.8410499 -0.28349996 -1.2568499 c -0.23624992 -0.3402 -0.6425998 -0.56700003 -1.10565 -0.56700003 c -0.61424994 0 -1.2473999 0.3402 -1.48365 0.96389997 c 0.3591001 -0.028349996 0.53865004 0.20790005 0.53865004 0.46305 c 0 0.40635002 -0.34965003 0.48195004 -0.47249997 0.48195004 c 0 0 -0.47250003 0 -0.47250003 -0.51030004 c 0 -0.8505 0.7749 -1.701 1.9089 -1.701 c 1.2096 0 2.2679996 0.89775 2.2679996 2.08845 Z "/></symbol></defs></svg></span> 在 <span role="math"><svg class="typst-frame" style="overflow: visible; width: 1.0243em; height: 0.683em;" viewBox="0 0 13.828050000000001 9.2205" width="13.828050000000001pt" height="9.2205pt" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" xmlns:h5="http://www.w3.org/1999/xhtml"><g><g class="typst-text" transform="matrix(1 0 0 -1 0 9.2205)"><use xlink:href="#g21E0DFDE6212E4409ACBE5D7BF48012D" x="0" y="0" fill="#000000" fill-rule="nonzero"/></g><g class="typst-text" transform="matrix(1 0 0 -1 7.695 12.555000000000001)"><use xlink:href="#g53818CE14381D89ADAFC63CFBFE9057F" x="0" y="0" fill="#000000" fill-rule="nonzero"/></g></g><defs id="glyph"><symbol id="g21E0DFDE6212E4409ACBE5D7BF48012D" overflow="visible"><path d="M 0 0m 7.6545 5.4945 c 0 0.32399988 -0.28349972 0.32399988 -0.53999996 0.32399988 h -4.5225 c -0.29699993 0 -0.80999994 0 -1.4039999 -0.6345 c -0.4725001 -0.52649975 -0.82350004 -1.1475 -0.82350004 -1.2149999 c 0 0 0 -0.13499999 0.16199997 -0.13499999 c 0.10800004 0 0.13500005 0.0539999 0.21600002 0.16199994 c 0.66150004 1.0395 1.4445 1.0395 1.7145 1.0395 h 0.7695 c -0.43199992 -1.6334999 -1.161 -3.267 -1.728 -4.4955 c -0.10799992 -0.20250002 -0.10799992 -0.22950003 -0.10799992 -0.324 c 0 -0.2565 0.21599996 -0.36450002 0.3915 -0.36450002 c 0.40499997 0 0.513 0.37800002 0.67499995 0.87750006 c 0.18900013 0.621 0.18900013 0.64800006 0.36450005 1.323 l 0.75600004 2.9835 h 1.5255 c -0.4454999 -1.9979999 -0.5669999 -2.5785 -0.5669999 -3.483 c 0 -0.20249999 0 -0.567 0.1079998 -1.026 c 0.13500023 -0.594 0.2835002 -0.675 0.48600006 -0.675 c 0.26999998 0 0.5535002 0.24300002 0.5535002 0.513 c 0 0.081 0 0.10800001 -0.08100033 0.29700002 c -0.3915 0.97200006 -0.3915 1.8495002 -0.3915 2.2275002 c 0 0.7154999 0.094500065 1.4444997 0.24300003 2.1464999 h 1.539 c 0.1755004 0 0.6615 0 0.6615 0.4590001 Z "/></symbol><symbol id="g53818CE14381D89ADAFC63CFBFE9057F" overflow="visible"><path d="M 0 0m 4.8761997 3.01455 c 0 1.0395 -0.12284994 1.7860501 -0.55754995 2.44755 c -0.29294968 0.4347 -0.87884974 0.8126998 -1.6348498 0.8126998 c -2.1924 0 -2.1924 -2.57985 -2.1924 -3.2602499 c 0 -0.68040013 0 -3.2035499 2.1924 -3.2035499 c 2.1923997 0 2.1923997 2.5231497 2.1923997 3.2035499 Z m -0.8599496 0.12284994 c 0 -0.69930005 0 -1.3986 -0.12285018 -1.9939499 c -0.18900013 -0.85995007 -0.83159995 -1.06785 -1.2096 -1.06785 c -0.4347 0 -1.01115 0.25515 -1.20015 1.0300499 c -0.13230002 0.55754995 -0.13230002 1.3324499 -0.13230002 2.03175 c 0 0.6898501 0 1.4080498 0.14174998 1.9277999 c 0.19845009 0.7465501 0.8032501 0.9450002 1.1907 0.9450002 c 0.5102999 0 1.0016999 -0.31185007 1.1717999 -0.85995007 c 0.15119982 -0.51030016 0.16065025 -1.1907 0.16065025 -2.01285 Z "/></symbol></defs></svg></span> 的基础上联合训练多机器人数据、Web 数据、高层语义预测与低层动作。训练样本混合图像、语言、目标检测、子任务预测和动作，使语义知识能够迁移到控制中。作者展示了模型在全新住宅内完成清理厨房、卧室等长时序灵巧任务的能力。它的重要性不只在模型规模，而在于说明异构数据与中间语义任务可以共同支撑端到端 VLA 的开放世界泛化；代价则是系统对数据覆盖、模型规模和推理基础设施的依赖很强。</p>