# iManip: Skill-Incremental Learning for Robotic Manipulation

**Authors:** Zexin Zheng, Jia-Feng Cai, Xiao-Ming Wu, Yi-Lin Wei, Yu-Ming Tang, Wei-Shi Zheng  
**arXiv:** [2503.07087](https://arxiv.org/abs/2503.07087)  
**Date:** 2025-03-10  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2503.07087v1/

## Key points

- Development of a generalist agent with adaptive multiple manipulation skills is a significant goal in robotics
- Paper titled "iManip: Skill-Incremental Learning for Robotic Manipulation" by Zexin Zheng et al. addresses this goal
- iManip framework proposed to overcome challenges faced by traditional incremental methods in robotic manipulation tasks
- Traditional methods often suffer from severe catastrophic forgetting due to overlooking key characteristics like temporality and action complexity
- iManip introduces temporal replay strategy to preserve old skills while learning new ones
- PerceiverIO component with adjustable weight introduced to adapt to new action primitives in novel skills
- Extensive experiments show that iManip framework excels in skill-incremental learning scenarios
- Authors plan to make codes for their environment and framework open-source for further research and development

## Summary

In the field of robotics, the development of a generalist agent with adaptive multiple manipulation skills has long been a significant goal. This goal is addressed in the paper titled "iManip: Skill-Incremental Learning for Robotic Manipulation" by authors Zexin Zheng, Jia-Feng Cai, Xiao-Ming Wu, Yi-Lin Wei, Yu-Ming Tang, and Wei-Shi Zheng. They delve into the crucial task of skill-incremental learning in robotic manipulation and propose an innovative framework called iManip to overcome challenges faced by traditional incremental methods. The authors first establish a skill-incremental environment using the RLBench benchmark and investigate how traditional methods perform in this setting. They discover that these methods often suffer from severe catastrophic forgetting due to overlooking key characteristics such as temporality and action complexity in robotic manipulation tasks. To address this issue, they introduce a temporal replay strategy in their iManip framework to preserve old skills while learning new ones. Additionally, the extendable PerceiverIO component is introduced featuring an action prompt with adjustable weight to adapt to new action primitives in novel skills. Through extensive experiments, the authors demonstrate that their iManip framework excels in skill-incremental learning scenarios. They plan to make the codes for their skill-incremental environment and framework open-source for further research and development in this area. Overall, their work represents a significant step forward in enabling robots to continuously acquire and refine manipulation skills without experiencing catastrophic forgetting.