Reviewer 21 of IROS 2026 submission 2076

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

The paper presents a crowd navigation framework that
explicitly models occlusion and agent interactions. The
method combines a VAE-based occlusion inference module with
attention-based human–human and human–robot interaction
learning, and trains a navigation policy using PPO.

Strengths
1. The problem is well-motivated. The paper addresses the
unrealistic full-observability assumption in crowd
navigation and highlights the importance of occlusion in
real environments.
2. The overall architecture is coherent. The combination of
occlusion inference and interaction modeling is intuitive
and aligns with the challenges of crowd navigation.
3. The interaction module distinguishes between human–
human and human–robot interactions, which is a reasonable
design choice for modeling heterogeneous dynamics.
4. The experimental results show improvements in success
rate and collision rate compared to baseline methods.
5. The ablation study indicates that combining occlusion
and interaction modeling improves performance over using
either component alone.

Weaknesses
1. The novelty is limited. The method mainly combines
existing ideas (VAE-based occlusion inference and
attention-based interaction modeling) rather than
introducing a fundamentally new approach.
2. The occlusion inference module (Eq. 2–3) is not well
analyzed. The choice of latent representation z and the use
of MSE reconstruction loss instead of likelihood-based
objectives are not justified, and it is unclear what
information z captures.
3. The reward function (Eq. 13) includes shaping terms,
but their influence on policy behavior is not analyzed, and
it is unclear whether they bias the policy toward specific
behaviors.
4. The simulation setup is simplified. Humans do not react
to the robot, which limits the realism of interaction
modeling and weakens the claim of improved social
navigation.
5. The occlusion modeling using partial OGMs is not clearly
linked to real sensor characteristics, and its realism is
unclear.
6. Important implementation details are missing, including
VAE architecture, latent dimension, and sensitivity to key
parameters.
7. The reported performance of key baselines (e.g.,
i-AW-CL and AttnGraph) differs significantly from their
original publications. For instance, i-AW-CL reports much
higher success rates and lower collision rates in its
original paper, while the values presented here are
noticeably lower. It is unclear whether this discrepancy is
due to differences in implementation, training protocols,
or evaluation settings. Since these baselines are central
to the claimed improvements, the authors should provide a
clear explanation; otherwise, the fairness and validity of
the comparison are questionable.


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

The video shows the performance of the method in both
simulated and real-world scenarios