Comments to author (Associate Editor)
=====================================

The problem framing is relevant and the three-module
architecture (VAE occlusion inference, dual-attention
interaction, PPO planning) is coherent and
well-implemented. Real-world deployment strengthens the
submission. However, revision must address four points
before acceptance can be recommended.
First, the reported performance of i-AW-CL and AttnGraph
differs substantially from their original publications. The
authors must explain this discrepancy explicitly —
differences in environment, sensor range, crowd density,
and human behavior model should be tabulated. Without this,
the claimed improvements are not verifiable.
Second, HEIGHT (Liu et al., T-ASE 2026), the direct
evolution of AttnGraph handling heterogeneous interactions
and constrained environments, should be added as a baseline
or explicitly discussed as out of scope with justification.
Third, the "Human-Robot Interaction" terminology is
inaccurate throughout: since all humans follow ORCA and do
not react to the robot, the cross-attention captures the
robot's one-sided response to human states, not genuine
bidirectional HRI. Please correct accordingly, as noted by
Reviewer 15.
Fourth, the VAE design is under-documented: the
architecture, latent dimension, pre-training dataset size,
and choice of MSE over likelihood-based loss are all
absent. At least one attention weight visualization should
be added to substantiate the interaction module's behavior.


----------------------------------------
Comments on Video Attachment:

Marginal — Video does not demonstrate the key
occlusion-avoidance scenarios central to the paper's
contribution.