Reviewer 15 of IROS 2026 submission 2076

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

Review of the Proposed Architecture:
The proposed work claims to be a novel formulation for safe
and efficient robot crowd navigation by simultaneously
addressing the challenges of severe sensor occlusions and
complex pedestrian interactions. The authors propose an
end-to-end framework consisting of three primary
components. First, an occlusion inference module utilizes a
Variational Autoencoder (VAE) to implicitly infer the
probabilistic states of hidden regions by processing a
sequence of historical, partially observed occupancy grid
maps. Second, an interaction learning module constructs a
spatial-temporal graph, employing a dual-attention
mechanism to separately model human-human interactions
(HHI) and human-robot interactions (HRI). Finally, the
inferred occlusion uncertainties and the aggregated
interaction features are fused and fed into an on-policy
Proximal Policy Optimization (PPO) reinforcement learning
algorithm to generate foresighted, collision-free
navigation actions.

Strengths: 
i)	The paper is generally well-written barring a few
terminological discrepancies. 
ii)	The dual-attention framework is technically sound
and has potential. I especially like how the attention
setup is decoupled to account for both human-human and
human-robot. The final context vector from this module
would make for a really good condition.
iii)	Coupling occlusion awareness with human context is
the natural progression from previous work PaS and
AttnGraph. 
iv)	The authors pre-train a VAE to infer occlusions.
During the RL training phase, the authors ignore the VAE
decoder and use the Encoder latent space representation as
conditioning. This is also consistent with the PaS paper
and the authors justify this design choice with superior
success rates through proper ablation.

The overall architecture is technically sound but not
without inconsistencies explained below.
Weaknesses:
i)	Limited Novelty: Looking at PaS Crowd Navigation,
and AttnGraph, the authors combined Occlusion prediction
and the Human Interaction from respective papers to their
planning algorithm. There seems to be limited technical
novelty with merely the usage of a different set of models
and parameters, contrary to the author’s claim.
ii)	Terminology for Human-Robot Interaction in the
Interaction module is incorrect because the humans are
controlled by ORCA and the interaction is only
one-sided(from the robot’s side). This is foresighted
planning and not true Human-Robot Interaction. The authors
need to correct this.
iii)	The authors completely omit the size and scope of
the pre-training dataset for the VAE-based occlusion
module, failing to mention the total number of frames,
episodes, or specific scenarios collected.

Question:
i)	Trajectory Tracking Discontinuity: The authors
mention in their Methodology section, how they extrapolate
the trajectories of each human in the environment to inform
their planning. This is perfectly consistent with the idea
of intention-aware planning. However, it is a discrepancy
with regards to the core premise of the paper, i.e,
occlusion-aware planning. If a human agent was previously
occluded by a physical barrier or a cluster of other
pedestrians and only becomes visible to the robot's sensors
at a particular time, the historical trajectory sequence
will be inherently incomplete. The paper fails to address
how the off-the-shelf trajectory extrapolator handles
occlusion and re-identification and preserves trajectory
continuity.

Review of the Results Section:
Strengths:
i)	The authors show considerable improvements over
many baselines, especially PaS and AttnGraph.

Weaknesses:
i)	The main drawback of the experimental results is
the fact that the authors trained and tested the robot in
the exact same 12m x 12m 2D simulation area. There are no
ablation studies or tests conducted in other simulated
environment geometries. There are considerable doubts as to
if the pipeline would generalize well to new environments.
No conclusive evidence has been presented in this regard. 
ii)	Authors show experiments with only 20 humans in the
scene. Although the crowd density looks high in simulation
and real-world tests modern SOTA baselines have proven to
be highly effective in scenarios with more than 50 dynamic
agents.
iii)	A crowded setting is expected to not just have
humans but also several static obstacles including walls.
Although not particular to the proposed research, it would
help if the authors also show qualitative examples of how
the planner handles static obstacles.  
iv)	The authors can strengthen their results by
comparing with recent crowd navigation baselines like
DenseCAvoid, and CrowdSurfer etc., especially the latter
because of their shared use of VAE-based formulation.

Future Improvements:

i)	The authors need to fix the Human-Robot-Interaction
Terminology throughout the paper. The proposed method is
just a foresighted planner that simultaneously takes into
account human and robot states. There is no interaction as
such since the humans are inherently controlled by ORCA.
ii)	The authors are requested provide test results
after training in multiple environments. Training in a
12x12 open space does not clarify if the model would
generalize to other environments. Similarly, no
qualitative/quantitative proof has been provided for the
planner also avoiding static obstacles, which are assumed
to be omnipresent in dense settings.
iii)	The authors are requested to show ablation studies
with more than 20 humans in different environments.


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

Video results are indeed showcasing the strengths of the
paper. Especially the qualitative improvements over
baselines