Intention-guided imitation learning methods under limited expert demonstration data

Abstract

Imitation Learning has achieved significant results in various fields, such as robot control, autonomous driving, and unmanned vessel decision-making. This technology aims to mimic human behavior in specific tasks by learning the mapping between states and actions, enabling agents to execute tasks based on demonstrations. However, these methods rely on the acquisition of high-quality demonstration data, facing challenges such as difficulties in collecting expert samples, high costs, and low efficiency in policy learning. Particularly under limited sample conditions, imitation learning can easily get trapped in local optima, leading to lower success rates and accuracy in decision-making. Researchers have used data augmentation and transfer learning to tackle limited data. However, in complex scenarios, these methods are less effective due to a lack of domain-specific knowledge, which affects the interpretability of the model. To address these challenges, we propose an Intention-guided Imitation Learning method under limited expert demonstration data (ITIL), which extracts deep intent features from a small number of samples to enhance the agent’s understanding of the scene and improve the accuracy of the mapping from states to actions during Imitation Learning. Specifically, the core method consists of three modules: (1) Semantic Enhancement Module, which extracts spatiotemporal feature maps from a small number of raw trajectories to enrich the semantic information of expert data; (2) Intention Expression Module, which constructs an intention tree network to establish connections between different levels, effectively expressing and capturing expert intent; (3) Strategy Generation Module, which integrates the outputs of the first two modules as input to form efficient decision-making, creating a closed-loop architecture of cognitive understanding-knowledge expression-decision optimization. Experimental results show that our model outperforms baseline methods in navigation, capture, and formation tasks, with an average success rate improvement of approximately +6 % compared to the baseline method (ValueDICE).