
ICRA 2026|新国立Bi-Adapt:基于语义对应的少样本双臂适配,实现跨类别泛化
Quick Take
Bi-Adapt is a novel framework for few-shot bimanual adaptation that leverages semantic correspondence to achieve cross-category generalization in robotic manipulation tasks, demonstrating high success rates even with limited data. The method integrates contact point selection to enhance bimanual coordination across five complex tasks and various object categories.
Key Points
- Introduces a unified framework for cross-category bimanual manipulation.
- Achieves effective generalization with few-shot and zero-shot learning.
- Demonstrates high success rates in simulations and real environments.
- Enhances bimanual coordination through contact point selection strategies.
- Evaluated across five complex tasks with various object categories.
Article Excerpt
From source RSS / original summary原文作者:新国立具身智能LinS Lab实验室主页:https://linsats. github. io/ Bi-Adapt: Few-Shot Bimanual Adaptation 1工作简介双臂操作(bimanual manipulation)是机器人完成复杂任务的关键能力,但现有方法通常依赖昂贵的数据采集与训练,并且难以泛化到未见类别的物体。 2我们提出了 Bi-Adapt,一个基于视觉基础模型(Vision Foundation Models)的双臂操作学习框架。 该方法通过引入语义对应(semantic correspondence),实现跨类别的 affordance 映射。 在仅使用极少量新类别数据进行适配的情况下,Bi-Adapt 仍能在未见类别物体上实现有效泛化(few-shot + zero-shot)。
3 核心贡献• 提出一个基于基础模型的统一框架,实现跨类别、跨任务的双臂操作泛化• 设计结合接触点选择(contact point selection)的 few-shot 适配策略,有效提升双臂协同能力• 在 5 类复杂任务与多种物体类别上进行系统评估(仿真 + 真实环境),在数据受限条件下仍取得高成功率 4 总结Bi-Adapt 通过语义对应实现跨类别的 affordance 迁移,并结合 few-shot 学习高效适配新类别,在有限交互数据下显著提升了双臂操作在未见物体上的泛化能力。 实验结果表明,该方法在 novel categories 上具备稳定且高效的适应性能。 项目主页https://biadapt-project. github. io/ 论文(arXiv)https://arxiv. org/abs/2602. 08425 代码开源https://github. com/isabella4444x/Biadapt 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.


