TacGNN
基于触觉与层级图神经网络的无视觉手内操作，重点看动作表示、图结构和真机证据。
Tactile in-hand manipulation without vision, focusing on action representation, graph structure and real-robot evidence.
论文阅读
把论文读成可以带走的证据、边界和方法。
好的论文阅读不是复述摘要，而是建立一条能回到图表、实验设置和失败边界的证据链。
A useful paper note is not a summary. It is an evidence trail that leads back to the figures, experiment setup and failure boundaries.
阅读方式
Reading method
证据 / 边界 / 代价 / 迁移
Evidence / Boundary / Cost / Transfer
把论文变成可以带走的方法，而不是一次性摘要。
Turn a paper into a reusable method rather than a one-off summary.
ReAct
把推理轨迹与行动交替结合起来，是理解工具调用型 Agent 的重要起点。
Interleaving reasoning traces with actions, and a useful starting point for tool-using agents.
RT-2
将视觉语言模型知识迁移到机器人动作，是讨论泛化、动作 token 与数据效率的入口。
Transferring vision-language knowledge into robot actions, with a focus on generalisation, action tokens and data efficiency.
Diffusion Policy
从动作分布、训练目标、推理频率和真机鲁棒性理解扩散策略的适用边界。
Understanding diffusion policies through action distributions, training objectives, inference frequency and real-robot robustness.
每篇论文只回答三个核心问题
Three questions for every paper
- 01
它改变了什么
What changed
是表示、规划、控制、数据、评测还是系统集成上的新能力。
A genuine advance in representation, planning, control, data, evaluation or integration.
- 02
证据够不够强
Is the evidence strong
结论是否建立在合理基线、重复实验、真机环境和清晰消融之上。
Whether the claim rests on fair baselines, repeated experiments, real hardware and clear ablations.
- 03
能否迁移
Can it transfer
机制能否进入自己的机器人或 Agent 系统，代价是什么。
Whether the mechanism can enter my own robotic or agent system, and at what cost.