ICRA 2026 | 神经衰减机制提升灵巧手精细抓握：双阶段深度预测学习TaSA框架，插入任务成功率翻倍
Quick Take
The TaSA framework, developed by Waseda University's robotics lab, introduces sensory attenuation in deep predictive learning, doubling task success rates for fine manipulation tasks like inserting pencil leads and coins. This two-phased approach enhances tactile recognition by filtering internal interference, significantly improving robotic dexterity.
Key Points
- TaSA framework integrates sensory attenuation to improve robotic manipulation.
- Achieves high tactile recognition despite self-touch interference.
- Two-phased learning: self-touch prediction and external interaction control.
- Demonstrated success in extreme fine tasks like inserting pencil leads.
- PCA analysis confirms clearer feature boundaries for real object contact.
Article Excerpt
From source RSS / original summary原文作者：研梦非凡人工智能原文链接：https://zhuanlan. zhihu. com/p/2031071527085528949 该研究由日本顶尖学府早稻田大学（Waseda University）的知名机器人实验室（菅野重树、尾形哲也团队）等联合提出。 arXiv 论文直达链接：https://arxiv. org/abs/2602. 05468 TaSA: Two-Phased Deep Predictive Learning of Tactile Sensory Attenuation for Improving In-Grasp Manipulation (ICRA 2026) 【亮点速览】引入人类“感觉衰减”机制：人类在抓握时能本能地忽略自己手指间的相互触碰，而专注于物体的触感。 但机器人多指灵巧手在操作时，手指间的碰撞（自我触碰）产生的触觉信号往往会淹没外部物体的信号。 TaSA框架首次在深度预测学习中引入了感觉衰减（Sensory Attenuation）机制。
解锁高难度精细操作：赋予了机械手极高的触觉辨识力，使其能够在存在大量自我触碰干扰的情况下，完成将自动铅笔芯插入笔筒、硬币投币、回形针夹纸等极端精细的任务。 【创新点】双阶段深度预测学习 (Two-Phased DPL)：第一阶段（自我触碰学习）：训练一个全连接网络（FCN），仅根据关节位置预测手指间自我触碰产生的触觉反馈；第二阶段（运动学习）：利用LSTM结合原始触觉输入和阶段一预测的自我触碰信号，过滤掉内部干扰，专注于外部物体的交互控制。 触觉特征空间净化：通过将自我预测作为基准进行剥离，算法有效缩小了由手指预紧力带来的“噪声方差”，使得真实物体接触的特征边界更加清晰（通过PCA分析证实）。 【成果】图2：TaSA（触觉感觉衰减）双阶段深度预测学习框架。 雷峰网
Reader Mode unavailable (could not extract clean content).
Want this in your inbox every morning?
Daily brief at your local 8am — bilingual EN/中文, free.
More from 雷峰网机器人
See more →
ICRA 2026 | 港中文高源、林天麟团队提出自发共适应策略：基于元学习赋能异构多机器人系统协同进化
The Hong Kong Polytechnic University team presented a meta-learning framework for heterogeneous multi-robot systems at ICRA 2026, enhancing adaptability in dynamic environments. Their approach improved task completion efficiency by 21% and reduced human burden by 39%, demonstrating effective human-robot interaction.
