Target-Driven Visual Navigation Using Causal Intervention

Target-driven visual navigation has gained significance and presents great potentials in scientific and industrial fields. However, how to achieve faster convergence and better generalization is a challenging problem. One of the most critical hurdles is the neglect of confounders, which often leads to spurious correlations. Confounders make it difficult to discover the real causality and therefore are taken into consideration in some other fields. In this paper, we introduce a Causal Intervention Visual Navigation (CIVN) method, based on deep reinforcement learning and causal inference. We propose to realize causal intervention in navigation via front-door adjustment as most confounders are unobservable. Specifically, CIVN is implemented by Target-Related Shortcut, which serves as an approximation of causal intervention. To eliminate the confounding effect, we adapt cross-sampling and strengthen the target information. It is worth mentioning that causal intervention is for the first time applied by us in solving the confounding effect in target-driven visual navigation. Navigation results on AI2-THOR demonstrate that CIVN converges faster and achieves better evaluation performance than prior arts. Moreover, the generalization for unknown targets and scenes is also improved.