Cross-Attention Guided Data Sharing for Knowledge Transfer in Robotic AI Systems

Abstract

Cross-robot transfer learning is crucial for building robotic AI systems that can generalize across diverse platforms and tasks by utilizing heterogeneous datasets. However, not all source samples are effective to improve the accuracy of the target AI task; while incompatible samples may lead to negative transfer and degrade model performance. This challenge is particularly in a connected robot fleet where robots differ in configurations, sensors, but are connected via Industrial Internet for sequential or parallel tasks. To address this, we propose Cross-Attention guided Proximal Policy Optimization (CAPPO), a reinforcement learning-based sample selection framework that adaptively identifies the most valuable source samples for a given target AI modeling task. Our method employs cross-attention mechanisms to capture fine-grained relevance between source and target samples, constructing informative state representations for a PPO-based selection policy. A task-driven reward function based on downstream performance improvement is created to enable the agent to learn efficient and adaptive selection strategies. Experimental results on a connected robotic fleet with different AI tasks show that our method consistently outperforms existing baselines under low-budget settings, demonstrating strong and robust knowledge transfer performance to train new robotic AI models.