Rethinking Latent Redundancy in Behavior Cloning: An Information Bottleneck Approach for Robot Manipulation

Behavior Cloning (BC) is a widely adopted visual imitation learning method in robot manipulation. Current BC approaches often enhance generalization by leveraging large datasets and incorporating additional visual and textual modalities to capture more diverse information. However, these methods overlook whether the learned representations contain redundant information and lack a solid theoretical foundation to guide the learning process. To address these limitations, we adopt an information-theoretic perspective and introduce mutual information to quantify and mitigate redundancy in latent representations. Building on this, we incorporate the Information Bottleneck (IB) principle into BC, which extends the idea of reducing redundancy by providing a structured framework for compressing irrelevant information while preserving task-relevant features. This work presents the first comprehensive study on redundancy in latent representations across various methods, backbones, and experimental settings, while extending the generalizability of the IB to BC. Extensive experiments and analyses on the CortexBench and LIBERO benchmarks show consistent performance improvements with IB across various settings, underscoring the importance of reducing input data redundancy and highlighting its practical value for real-world applications.

How can behavior cloning (BC) be improved by compressing its learned representations?In BC, observations are encoded into latent representations before being mapped to actions. We ask whether these representations retain unnecessary information. To investigate this, we apply the Information Bottleneck (IB) principle to reduce the mutual information between the input and the latent space, aiming to remove task-irrelevant features.Our findings show that latent representations in standard BC are not optimal.By compressing input information, we suppress irrelevant information and improve performance in robotic manipulation tasks. This suggests that focusing on essential information helps BC generalize better, even without adding model complexity.This study highlights a fundamental trade-off between compression and generalization.Too much information can hurt performance, while targeted compression can enhance it. Our work shows that information-theoretic tools like IB offer a promising direction for designing more efficient and robust imitation learning algorithms.