2025具身智能论文风向标：看清华TEA Lab与小红书薯友如何票选“年度最佳”论文
当全球追逐算力军备竞赛时，一小群学者却在悄悄定义“品味”。
最近，清华大学TEA Lab评选出了一份独特的 “2025具身智能年度论文” 清单。
在小红书博主@许华哲Harry的分享中，这份清单没有罗列冰冷的数据，而是颁出了 “好莱坞Demo奖”、“奥卡姆剃刀奖”、“最佳开源奖” 等趣味奖项。
他们试图回答：在算力之外，什么才是真正值得品味的智能？
本文已获得许华哲老师授权分享，论文清单如下👇
01 对多任务灵巧操作的大型行为模型的仔细研究
原文标题：
“A Careful Examination of Large Behavior Models for Multitask Dexterous Manipulation”
原文链接：
https://arxiv.org/abs/2507.05331
"This shows impressive results with in-depth analysis. To the field, this is morelike an ankor point where everyone should start to believe in large models ratherthan a ton of smaller models. I like how TRI do things that are both scalable anddetailed. "
02 π^{*}_{0.6}：一个从经验中学习的VLA
原文标题：
“π^{*}_{0.6}: a VLA That Learns From Experience”
原文链接：
https://arxiv.org/abs/2511.14759
"High success rate for long-horizon dextrous manipulation tasks using iterativeoffline reinforcement learning"
03 关于噪声的大惊小怪：打破生成式机器人控制的神话
原文标题：
“Much Ado About Noising: Dispelling the Myths of Generative Robotic Control”
原文链接：
https://arxiv.org/abs/2512.01809
"A comprehensive evaluation of popular generative control policies (GCPs) oncommon behavior cloning (BC) benchmarks, the paper suggest that thedistribution-fitting component of GCPs is less salient than commonly believed,and point toward new design spaces focusing solely on control performance.'
04 DexUMI：将人手作为灵巧操作的通用操控界面
原文标题：
“DexUMI: Using Human Hand as the Universal Manipulation Interface for Dexterous Manipulation”
原文链接：
https://arxiv.org/abs/2505.21864
"Scaling human hand data to real dexterous hand hardware without retargetingissues. Drastically improve data collection efficiency with relatively low-cost andhigh-quality, though inpainting is still needed for visual alignment.'
05 一天学会一千项任务
原文标题：
“Learning a Thousand Tasks in a Day”
原文链接：
https://arxiv.org/abs/2511.10110
"Demonstrating the power of efficient imitation learning in the era of scalingup.'
06 π_{0.5}：一个具有开放世界推广的视觉-语言-行动模型
原文标题：
“π_{0.5}: a Vision-Language-Action Model with Open-World Generalization”
原文链接：
https://arxiv.org/abs/2504.16054
"The first highly generalizable VL A model."
07 VLASH：通过面向未来状态的异步推理实现实时VLA
原文标题：
“VLASH: Real-Time VLAs via Future-State-Aware Asynchronous Inference”
原文链接：
https://arxiv.org/abs/2512.01031
"Real-time chunking is a fundamental yet critical component of VL,A models,effectively serving as the controller: Investigating how 1o derive such a robustcontroller presents a valuable research problem.
08 VGGT：基于视觉几何的变换器
原文标题：
“VGGT: Visual Geometry Grounded Transformer”
原文链接：
https://arxiv.org/abs/2503.11651
"Very impactful on 3d representation learning. I think that there could be some project to do using vggt for robotics vision issues"
09 DiffusionNFT：在线扩散 正向过程强化
原文标题：
“DiffusionNFT: Online Diffusion#9Reinforcement with Forward Process”
原文链接：
https://arxiv.org/abs/2509.16117
"This paper clarifies the role of the old policy in diffusion, bypasses the need toexplicitly define an intractable policy density pfor policy gradients, andstabilizes DPO-style negative optimization by anchoring the negative target to 2old-v-. Moreover, it provides a clear interpretation of what Flow-GRPO is trulyoptimizing and demonstrates how CFG-like effects can emerge without explicitclassifier-free guidance."
10 手眼自主配送：学习类人导航、运动和伸手能力
原文标题：
“Hand-Eye Autonomous Delivery:Learning Humanoid Navigation,Locomotion and Reaching”
原文链接：
https://arxiv.org/abs/2508.03068
Autonomous visual navigation + eyes /hands tracking: one meaningful steptoward general-purpose loco-manipulation.'
