Reviewer 8 of IROS 2026 submission 2076

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

Navigating dense crowds remains a fundamental challenge for
autonomous robots, as most existing frameworks assume
complete visibility of human dynamics, which is often
compromised by frequent occlusions in real-world settings.
To address this, the authors propose a novel reinforcement
learning framework that simultaneously models occluded
regions and complex human-robot interactions. The method
utilizes a variational autoencoder (VAE) as an occlusion
inference module to implicitly estimate the states of
hidden areas, paired with an attention-based interaction
learning module to capture the subtle social dependencies
within the crowd. These representations are fused and
trained using an on-policy reinforcement learning algorithm
to generate navigation policies that prioritize both safety
and efficiency. Benchmark evaluations against
representative baseline methods demonstrate that this
simultaneous consideration of occlusion and interaction
leads to superior collision avoidance performance and
smoother trajectories in high-density populated
environments.

Pros:
(1) The integration of both occlusion inference and
interaction learning addresses two of the most critical and
often overlooked practical constraints in crowd navigation.
(2) Using a VAE to "implicitly infer" occluded states is a
technically sound way to handle the high uncertainty
inherent in hidden environmental regions.
(3) The use of an attention mechanism to capture
dependencies ensures the robot can prioritize relevant
interactions rather than treating all nearby humans with
uniform weight.

Cons:
(1) The simultaneous execution of a VAE inference and an
attention-based RL policy may introduce significant
computational overhead for robots with limited onboard
processing.
(2) Relying on "implicit" inference for occlusions may lead
to unpredictable edge-case behaviors compared to frameworks
with explicit geometric safety bounds.
(3) While the policy is trained using on-policy RL, the
manuscript lacks a detailed discussion on the specific
noise models used to ensure the VAE generalizes from
simulation to real-world sensor noise.
(4) The current evaluation focuses on "representative
approaches" but does not deeply explore how the policy
handles heterogeneous crowds with diverse movement
behaviors (e.g., children or fast-moving cyclists).

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

N/A