Reviewer 19 of IROS 2026 submission 2076

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

This paper proposes a navigation framework designed for
robot navigation in dense crowd environments.The robot is
equipped with a nonglobal,limitedfieldofview sensor and
must navigate autonomously under partial observability.The
authors first introduce an occlusion inference module based
on a variational autoencoder(VAE),which leverages a
sequence of historical partially observed occupancy grid
maps(OGMs)to predict the state of occluded regions,thereby
reconstructing the global map and outputting it in the form
of latent variables.Meanwhile,they employ selfattention
and crossattention networks to extract humanhuman
interaction(HHI)and humanrobot
interaction(HRI)features,respectively.Finally,the latent
variables of the global map from the VAE and the
humanrobot interaction features are fed together into a
hierarchical reinforcement learning(HRL)framework,enabling
the robot to autonomously perform multitask
decisionmaking,including obstacle avoidance,crowd dynamics
perception,and goaldirected navigation.To validate the
effectiveness of the proposed method,the authors conduct
ablation studies on a simulation platform that explicitly
models occlusions,perform quantitative comparisons with
existing methods,and also deploy and verify the approach on
a real robot platform.

Strengths:
-Technical completeness:The work simultaneously addresses
occlusion reasoning and interaction modeling.The endtoend
framework—VAEbased latent variable extraction for
occlusion,dualattention networks for HHI and HRI,and PPO
for policy learning—is clearly structured and
wellintegrated,demonstrating strong technical
completeness.
-Thorough ablation study:The authors systematically
validate each component through modulelevel ablations and
a comparison between the latent variable and reconstructed
features.The results show that combining both occlusion and
interaction modules raises the success rate to 0.95 while
reducing the collision rate to 0.05,providing solid
experimental support for the key design choices.
-Clear and wellpresented:The paper effectively combines
figures(overview,OGM conversion,scenario comparisons)with a
stepbystep derivation of formulas.The terminology is
consistent,and the experimental
comparisons(e.g.,Table I)are easy to
follow.Overall,the presentation is logical,readable,and
facilitates reproducibility.

Weaknesses and Concerns:
-Insufficient realworld evaluation of occlusion handling
The supplementary video does not clearly demonstrate how
the robot responds to occluded pedestrians.The authors
should modify both simulation and realrobot experiments by
adding test scenarios in which a previously occluded target
suddenly appears and blocks the robot’s planned path.This
would allow proper assessment of the robot’s ability to
anticipate and avoid such risks.Currently,the study lacks a
quantitative evaluation of occlusion reasoning performance.
-Lack of intuitive visualization for the attention
mechanism
The ablation study shows that removing the attention module
leads to a significant performance drop(e.g.,a 23%reduction
in success rate).However,the paper does not provide a
direct evaluation metric for how the attention layers
contribute to the navigation task.The authors should
propose evaluation criteria and testing methods that
illustrate the actual role of the attention
mechanism,making its effects more understandable.
-Insufficient description of baseline methods
The paper compares several existing algorithms but does not
introduce them in the text.The authors should briefly
describe the core strategy of each baseline,as well as the
key differences from the proposed method,either in the
experiments section or in the related work section.This
would enhance the completeness and reference value of the
comparative analysis.

Minor Suggestions:
-Page 1:"trained in simulator"should be"trained in a
simulator"or"trained in simulators"
-Page 2:"future evolution of neighboring agent"should
be"future evolution of neighboring agents"
-Page 3:"there often occurs occlusion phenomenon"should
be"there often occurs an occlusion phenomenon"
-Page 4:"allows...to allocate...and thereby
improving"should be"allows...to allocate...and thereby
improve"or"allows...to allocate...,thereby improving"
-Page 8:"diffusion model"should be"diffusion models"

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

The video is concise and intuitive, but insufficient. It
should more clearly demonstrate how occluded obstacles
impact robot navigation, and how the robot handles such
occluded situations. Additional examples illustrating the
robot's ability to anticipate and avoid collisions with
occluded pedestrians would help validate the method's key
contribution.