iManip: Skill-Incremental Learning for Robotic Manipulation

TL;DR

This work tackles skill-incremental learning for robotic manipulation by identifying temporal dynamics and action-primitives expansion as core challenges that cause catastrophic forgetting in traditional baselines. It introduces iManip, combining a Temporal Replay Strategy with an Extendable PerceiverIO that uses skill-specific action prompts and extendable weights to learn new actions while preserving prior skills, aided by knowledge distillation. Across RLBench simulations and real-world robot experiments, iManip demonstrates superior retention of old skills, improved adaptation to new ones, and efficient fine-tuning. The approach advances lifelong robotic manipulation by enabling continual skill acquisition with open-source resources for the community.

Abstract

The development of a generalist agent with adaptive multiple manipulation skills has been a long-standing goal in the robotics community. In this paper, we explore a crucial task, skill-incremental learning, in robotic manipulation, which is to endow the robots with the ability to learn new manipulation skills based on the previous learned knowledge without re-training. First, we build a skill-incremental environment based on the RLBench benchmark, and explore how traditional incremental methods perform in this setting. We find that they suffer from severe catastrophic forgetting due to the previous methods on classification overlooking the characteristics of temporality and action complexity in robotic manipulation tasks. Towards this end, we propose an incremental Manip}ulation framework, termed iManip, to mitigate the above issues. We firstly design a temporal replay strategy to maintain the integrity of old skills when learning new skill. Moreover, we propose the extendable PerceiverIO, consisting of an action prompt with extendable weight to adapt to new action primitives in new skill. Extensive experiments show that our framework performs well in Skill-Incremental Learning. Codes of the skill-incremental environment with our framework will be open-source.

Figures (7)

• Figure 1: (a) An overview of our skill-incremental learning for robotic manipulation that requires the agent to learn skill sequences over time. (b) A comparison of model performance between the traditional incremental baseline (TIB) and our iManip.
• Figure 2: Overview of robotic incremental learning. Previous works focus on incremental abilities in new objects, goals, or spatial positions, where different tasks may share the same skill. The iManip focuses on skill-incremental learning which better captures the true adaptability and flexibility required for real-world robotic learning.
• Figure 3: The overall framework of iManip, which primarily consists of a temporal replay strategy to store the samples with the farthest distance entropy for each keyframe of old demonstrations and an extendable PerceiverIO consisting of action prompts with extendable weights to adapt to new action primitives.
• Figure 4: Average success rate on different replay methods.
• Figure 5: Visualization of skill-specific action prompts by Grad-CAM as the agent executes two manipulation skills: close jar (the first row) and slide block (the second row).