Reviewer 2 of IROS 2026 submission 2076

Comments to the author
======================

This paper addresses an important problem in crowd
navigation by jointly considering occlusion and
interaction. The proposed framework integrates a VAE-based
occlusion inference module with attention-based interaction
modeling and RL planning. Experimental results show
consistent improvements over several baselines. 

However, the overall novelty is somewhat limited. The
method mainly combines existing components (VAE, attention
mechanisms, PPO) without introducing fundamentally new
techniques. 

The experimental evaluation is generally solid, including
ablation studies and a real-world deployment. However, the
paper lacks cross-scenario validation and provides limited
evaluation of generalization. While the ablation experiment
showing that latent features outperform reconstructed
features is well designed, the analysis of each module and
their interactions is insufficient, remaining at a
phenomenological level and failing to substantiate the
optimality of the proposed design.

From a presentation perspective, the system framework
diagram is overly simplified, which affects the readability
of the method. In addition, the demonstration video is
rather brief and does not fully showcase the robustness and
advantages of the proposed method in diverse scenarios.

Comments on the Video Attachment
================================

The video shows the experiment but is not sufficient to
demonstrate the advantages of the proposed method.