← Back ICRA 2026

Bi-Adapt: Few-Shot Bimanual Adaptation for Novel Categories of 3D Objects Via Semantic Correspondence

Jinxian Zhou, Ruihai Wu, Yiwei Liu, Checheng Yu, Xunzhe Zhou, Yiwen Hou, Licheng Zhong, Lin Shao

PDF

AI summary

Key figure (auto-extracted from paper)

Bi-Adapt enables robots to efficiently generalize complex bimanual manipulation to novel 3D objects using only a few interactions by leveraging foundation model-guided semantic correspondence.

Bimanual manipulation Few-shot learning Semantic correspondence Affordance transfer Foundation models Robotic adaptation

Problem

Current bimanual manipulation methods depend on extensive data collection and training, failing to efficiently adapt to unseen objects across different categories.

Approach

The framework learns manipulation priors on a supporting set, maps contact points to novel objects using a vision foundation model, and fine-tunes the policy through minimal few-shot interactions.

Key results

Achieves 59–70% success rates across five bimanual tasks on unseen novel objects
Outperforms existing baselines in cross-category affordance generalization
Introduces a foundation-model-guided semantic correspondence mapping for contact point transfer
Demonstrates efficient policy adaptation requiring only minimal interaction data

Why it matters

It offers a scalable, data-efficient solution for robots to handle diverse bimanual tasks on unfamiliar objects, advancing practical deployment in human-centered environments.

Abstract

Bimanual manipulation is imperative yet challeng- ing for robots to execute complex tasks, requiring coordinated collaboration between two arms. However, existing methods for bimanual manipulation often rely on costly data collection and training, struggling to generalize to unseen objects in novel categories efficiently. In this paper, we present Bi-Adapt, a novel framework designed for efficient generalization for bimanual manipulation via semantic correspondence. Bi-Adapt achieves cross-category affordance mapping by leveraging the strong capability of vision foundation models. Fine-tuning with restricted data on novel categories, Bi-Adapt exhibits notable generalization to out-of-category objects in a zero-shot manner. Extensive experiments conducted in both simulation and real- world environments validate the effectiveness of our approach and demonstrate its high efficiency, achieving a high success rate on different benchmark tasks across novel categories with limited data. Project website: https://biadapt-project. github.io/

Index terms

Bimanual Manipulation Visual Learning Learning from Demonstration