# Occupancy-History Conditioning for Crowd Navigation under Occlusion **Editorial working draft v2 — revised after two review rounds for a first RA-L submission, 14 October 2026.** This draft preserves the experiments and numerical results reported in `IROS26_2076_MS.pdf`. It rewrites their interpretation; it does not report new experiments. Bracketed **AUTHOR CHECK** notes identify facts that cannot be established from the supplied manuscript. Resolve them from the implementation and existing records before submission. Reference keys are descriptive placeholders for the final bibliography. This is replacement manuscript text, not a submission-ready typeset paper. ## Abstract Occlusion makes crowd navigation depend on an incomplete set of pedestrian observations. We study two representation choices for this setting: combining occupancy history with detected pedestrian motion, and supplying occupancy features directly to the policy. A variational encoder is pretrained with complete local occupancy maps as supervision for sequences of partial maps. Its latent features are concatenated with pedestrian motion context and robot state, then supplied to a recurrent policy trained with proximal policy optimization. Under the reported simulation protocol with finite-range sensing and pedestrian occlusion, the combined configuration reports 95% success and 5% collision, compared with 87% success and 12% collision for the motion-only configuration. The occupancy-only configuration reports 72% success. A separate comparison reports 0% timeout with direct latent conditioning and 16% with decoded-and-reencoded occupancy features. A physical-robot demonstration illustrates execution with onboard perception. These results support the combined representation and direct latent interface in the evaluated architecture, while leaving the source and generality of the observed gains open. ## I. Introduction When pedestrians occlude one another, a robot's current detections provide an incomplete description of the surrounding space. A pedestrian absent from the current observation may still affect the robot's next action. Finite sensing range and obstruction within that range produce different sources of missing observations, both of which matter when navigating toward a goal. This work focuses on navigation under a specified model of finite-range sensing and pedestrian occlusion. Existing approaches offer two relevant ways to represent the available observations. Occupancy-based methods organize local observations spatially and can learn features from partial map histories; PaS already incorporates such latent features into a navigation policy [PaS]. Pedestrian-based methods retain individual motion information and aggregate it according to the robot state; AttnGraph combines trajectory prediction, attention, and recurrent policy learning [AttnGraph]. In our design, the motion branch uses detected pedestrian tracks and their predicted motion, while the occupancy branch learns from partial map histories using complete current local occupancy as a training target. At execution, both branches depend on partial observations. This difference in representation and supervision motivates our question: **does adding occupancy-history conditioning improve navigation outcomes when a policy already receives contextualized pedestrian motion features?** We investigate this question with a policy that combines the two representations. An occupancy history encoder is pretrained to reconstruct complete local maps from sequences of partial maps. A pretrained trajectory predictor supplies motion estimates for detected pedestrians; a motion encoder contextualizes pedestrian features and aggregates them using a query derived from the robot state. The occupancy latent, motion features, and robot state are concatenated and processed by a recurrent policy. We also examine a second design choice within this architecture: whether the occupancy information is supplied as an encoder latent or through a decoded map followed by a CNN. The approach builds on occupancy-based and attention-based navigation; neither concatenation nor direct latent navigation is claimed as a new general principle. The experiments examine the effect of visibility, the use of individual feature branches, and the representation supplied to the policy. The combined configuration reports 95% success, compared with 87% for the motion-only configuration and 75% for the decoded-and-reencoded configuration. These comparisons have different outcome profiles: the motion-only configuration reports 12% collision and 1% timeout, while the decoded-and-reencoded configuration reports 9% collision and 16% timeout; the combined direct-latent configuration reports 5% collision and 0% timeout. These observations motivate reporting collision and timeout separately, without assigning them to a unique causal mechanism. We also compare existing navigation methods and demonstrate execution on a physical robot. The evaluation concerns randomly sampled test episodes from the specified simulation setting; it does not establish performance across unseen scene geometries or arbitrary sensor failures. Our contributions are: 1. A recurrent navigation architecture that combines occupancy-supervised features of partial map histories, contextualized pedestrian motion, and robot state under finite-range sensing and simulated pedestrian occlusion. 2. An evaluation of two design choices within this architecture: combining occupancy and motion representations, and supplying occupancy features as a direct latent rather than through the tested decode-and-reencode interface. We report success, collision, and timeout separately, alongside baseline comparisons and a physical-robot demonstration. **AUTHOR CHECK — INTRODUCTION:** Confirm the ablation configurations and their evaluation sets before retaining the quantitative comparison as the central empirical claim. Do not describe the motion-only variant as a reproduction of AttnGraph unless their implementations and training protocols are shown to match. ## II. Related Work ### A. Occupancy Representations under Partial Observation PaS learns navigation-relevant features from partial occupancy histories using complete occupancy information during training, and supplies latent features directly to a recurrent policy [PaS]. Its training uses a complete-map teacher representation and a latent matching objective; our manuscript describes direct reconstruction pretraining from partial histories. PaS also observes that optimizing map reconstruction can conflict with navigation performance. Thus, neither occupancy-history encoding nor direct latent conditioning is introduced here as a new principle. SCOPE studies stochastic prediction of future occupancy and its use in navigation [SCOPE], whereas the supervision target in this work is the current complete local occupancy map. We study the joint use of occupancy-history features and an explicit pedestrian motion representation. ### B. Pedestrian Motion Representations for Navigation DS-RNN and AttnGraph use recurrent representations of surrounding pedestrians for navigation [DSRNN, AttnGraph]. AttnGraph incorporates future trajectory predictions and attention-based aggregation. HEIGHT extends the representation to heterogeneous scene elements, including static obstacles, and handles visibility through masking [HEIGHT]. These methods should not be characterized as uniformly assuming complete observations. Our motion branch follows the established sequence of per-pedestrian context encoding and robot-conditioned aggregation. We examine the additional use of an occupancy-supervised history representation in an open crowd setting. HEIGHT remains a closely related method; the present evaluation does not compare performance in its constrained environments or establish superiority over it. EgoTraj-Bench identifies trajectory-history corruption arising from first-person perception, including occlusion, identity switches, and tracking drift [EgoTraj]. Its paired noisy-history and clean-trajectory evaluation distinguishes realistic perception errors from idealized missing observations. Our simulator models visibility loss geometrically; it does not reproduce the full set of perception artifacts studied by that benchmark. The interface between detection, track history, and trajectory prediction must therefore be specified explicitly. More recently, iCrowdNav combines temporal BEV features from RGB-D images with pose-derived pedestrian features for policy learning [iCrowdNav]. Its architecture also concatenates scene and pedestrian representations. Our occupancy branch instead uses complete current local occupancy as supervision for partial map histories, while our pedestrian branch encodes current and predicted planar positions. These are differences in inputs and training formulation; the generic combination of scene and pedestrian features is not claimed as a new principle here. ### C. Navigation Guided by Pedestrian Motion People as Planners selects suitable pedestrians and local subgoals so that a base planner can use their movement as navigation guidance [PeoplePlanners]. HiCrowd learns a follow point for crowd-flow alignment and executes it with model predictive control [HiCrowd]. These methods make pedestrian motion useful through explicit guidance decisions. Our method instead supplies occupancy-history and pedestrian-motion features to a velocity policy. The present experiments evaluate navigation outcomes; they do not demonstrate leader selection, crowd-flow alignment, or a formal safety guarantee. **AUTHOR CHECK — RELATED WORK:** Integrate the original i-AW-CL citation and describe only verified method details. Its full experimental protocol was not available in the supplied material. The iCrowdNav comparison above was checked against Sections III-B and III-C of its arXiv v1; it does not imply a common evaluation protocol or a performance comparison. ## III. Method ### A. Observations and Policy Inputs The robot navigates among pedestrians using its proprioceptive state and partial observations. Let \(w_t\) denote its position, velocity, goal, heading, maximum-speed parameter, and radius. Let \(\mathcal D_t\) be the set of detected pedestrians at time \(t\), with \(n_t=|\mathcal D_t|\). The complete simulator state is used for training supervision and reward calculation, while the policy receives features computed from the available observations. For each detected pedestrian \(i\), a pretrained trajectory predictor [Predictor] uses its available historical positions to obtain future position estimates \(\hat x_i^{t+1:t+M}\). These estimates are model outputs, not ground-truth future observations. The motion-encoder input below follows the current-position-plus-prediction description in the original Section III-D, subject to the implementation check below. We use \(O_t\) for a partial occupancy grid, reserving separate notation for the complete observation and for recurrent hidden states. **AUTHOR CHECK — OBSERVATION INTERFACE:** Report \(K\), \(M\), sampling interval, predictor checkpoint and training data. Verify the actual tensor entering the motion MLP: the original Section III-A includes historical positions in the pedestrian representation, while Section III-D describes current positions and predicted futures. Confirm whether history feeds only the predictor or also the motion encoder, and align the equation, text, and diagram accordingly. Specify the implemented treatment of missing track samples, newly detected pedestrians, identity changes, reappearing tracks, and an empty detection set. Do not insert an interpolation, identity-recovery, or masking rule that was not used in the reported experiments. Check that simulator IDs and hidden positions are not used to populate unavailable history samples. ### B. Occupancy History Encoding Each partial local map encodes free, occluded, and occupied cells using values 0, 0.5, and 1, respectively. Here 0.5 is an encoding for unobserved space, not an empirically calibrated occupancy probability. The complete local target map \(G_t\) contains the occupancy of observed and occluded pedestrians within the specified local sensing region. It is not a global map of the environment. The encoder takes the history \(O_{t-\tau+1:t}\) and produces a latent distribution \(q_\phi(z\mid O_{t-\tau+1:t})\). Writing the reconstruction objective in operational form, the loss reported in the original manuscript is \[ \mathcal L_{\mathrm{rep}} =\mathbb E_{z\sim q_\phi}\! \left[\lVert G_t-\hat G_t\rVert_2^2\right] +D_{\mathrm{KL}}\!\left(q_\phi(z\mid O_{t-\tau+1:t})\,\|\,\mathcal N(0,I)\right). \] The reconstruction term encourages the representation to retain information useful for predicting the complete local occupancy target, and the KL term regularizes the latent distribution. A squared reconstruction loss can be related to a fixed-variance Gaussian observation model, subject to its scaling and the decoder parameterization. This interpretation does not establish calibrated occupancy probabilities or a calibrated uncertainty representation for the policy. During navigation, the policy receives an encoder-derived latent feature \(z_t\), without using a decoded map as its navigation input. This follows the direct-latent design used in prior occupancy-based navigation [PaS]. Our representation comparison evaluates it against decoding and reencoding the occupancy output within the present architecture. **AUTHOR CHECK — OCCUPANCY ENCODER:** Supply the grid size and resolution, coordinate frame and history alignment, \(\tau\), network layers, latent dimension, pretraining dataset size and split, optimizer, and actual loss normalization/weights. Confirm whether the decoder also receives the observation history; the original conditional generative equation and the diagram do not settle this. State whether the policy uses a latent sample, a mean, or another feature, and whether the encoder is frozen or updated during PPO. Do not describe the entire pipeline as end-to-end without verifying the gradient paths. ### C. Pedestrian Motion Context For each detected pedestrian, an MLP embeds its current position and predicted future positions: \[ e_i^t=\mathrm{MLP}\big([x_i^t,\hat x_i^{t+1:t+M}]\big), \qquad E_t=[e_1^t;\ldots;e_{n_t}^t]. \] A self-attention operation contextualizes these embeddings: \[ Q_c=\Phi_Q(E_t),\quad K_c=\Phi_K(E_t),\quad V_c=\Phi_V(E_t), \qquad C_t=\operatorname{softmax}\left(\frac{Q_cK_c^\top}{\sqrt{d_c}}\right)V_c. \] It produces one context feature per pedestrian. A second operation aggregates these features using a query derived from the robot state: \[ q_p=\Psi_Q(w_t),\quad K_p=\Psi_K(C_t),\quad V_p=\Psi_V(C_t), \qquad c_t=\operatorname{softmax}\left(\frac{q_pK_p^\top}{\sqrt{d_p}}\right)V_p. \] The first operation produces \(n_t\) contextualized features; the second produces one robot-state-conditioned summary. The attention coefficients specify feature aggregation within the network. They are not treated as measured collision probabilities or estimates of pedestrians' causal responses to the robot. Let \(f_t^{\mathrm{motion}}\) denote the motion branch output supplied to the policy. **AUTHOR CHECK — RECURRENCE:** The original Section III-D describes a temporal GRU, while Section III-E and Figure 2 describe recurrence after fusion. Confirm whether there is one GRU or two. If only the fused-policy GRU exists, set \(f_t^{\mathrm{motion}}=c_t\) and remove the earlier temporal-GRU statement. If a separate motion GRU exists, add its actual update equation and show both hidden states in the figure. This draft deliberately leaves that implementation choice unresolved. ### D. Recurrent Navigation Policy The policy combines the robot state, motion features, and occupancy latent by concatenation: \[ h_t=\mathrm{GRU}\big(h_{t-1},[w_t,f_t^{\mathrm{motion}},z_t]\big), \qquad a_t\sim\pi_\theta(\cdot\mid h_t). \] The actor produces a planar velocity command and the critic estimates the return. PPO [PPO] is used to optimize the policy. The complete local occupancy map is a training target; it is not a deployment input. The recurrent feature is a learned policy representation, without a claim of Bayesian filtering or decision sufficiency. We retain the original piecewise reward: \[ r_t=\begin{cases} 10,&d_t^{\mathrm{goal}}