VP-EVT: Active Visibility Preservation for Embodied Visual Tracking

Yanci Wen, Tianle Zeng, Jianwei Peng, Weixi Situ, Hanjing Ye, Hong Zhang

Oct 15, 2026arXiv:2610.21999v1

cs.RO

Scorecard· 16/16

5.5/10 impact

Well-framed active visibility preservation with large in-benchmark gains, a useful benchmark and thorough ablations, but a self-designed evaluation, thin baselines, and narrow subfield reach limit impact.

Abstract

Embodied Visual Tracking (EVT) is an essential capability for intelligent robots in human-robot collaboration, requiring an agent to follow a designated target person in dynamic environments relying solely on egocentric visual perception. However, persistent pedestrian occlusions in dynamic scenes pose critical challenges to existing EVT methods. Lacking proactive visibility preservation, reactive tracking methods frequently experience prolonged target loss under severe occlusions. Therefore, we propose VP-EVT, a framework that treats target visibility as an explicit objective of motion control rather than a by-product of following. Acting on this objective requires understanding both what obstructs the target and how robot motion can change that obstruction. VP-EVT therefore uses language-guided perception to extract separate target and pedestrian observations. When occlusion makes current observations incomplete, VP-EVT combines visual history with ego-motion to maintain estimates of their relative positions and motion. Using these estimates and the current following waypoints, VP-EVT predicts whether continued following will sustain the obstruction. When the predicted risk is low, the original following motion is retained; otherwise, a recurrent controller combines updated estimates with maneuver context from preceding decisions to coordinate viewpoint adjustment and the return to stable following. Moreover, we introduce OccEVT-Bench for closed-loop evaluation of persistent occlusion and collect expert demonstrations for training. On its Risk split, VP-EVT improves success rate by 121.4% relative to a competitive baseline and reduces visibility deficit and longest target-loss duration by 78.8% and 74.4%, respectively.

AI Impact Assessments

(1 model)

Core Contribution

VP-EVT reframes embodied visual tracking (EVT) under pedestrian occlusion. Target visibility becomes an explicit motion-control objective rather than a by-product of following. The system is a modular add-on to a frozen follower (OmTrackVLA) with three components:

Language-conditioned grounding that produces separate target and pedestrian observations.

Gaussian temporal beliefs, filtered with ego-motion, that persist through occlusion.

A plan-conditioned risk predictor that checks whether the current following waypoints will sustain the occlusion. A GRU controller then runs a viewpoint-adjusting maneuver and hands control back to the follower.

The paper also contributes OccEVT-Bench, a 400-episode closed-loop benchmark on Habitat-GS with 3DGS scenes. It has Risk and Safe splits and metrics for visibility deficit, longest target loss, and physical-occlusion versus out-of-FOV loss attribution. In addition, it releases OccEVT-Expert, 20k expert demonstration episodes.

The conceptual shift from reactive recovery to anticipatory visibility preservation is sensible and well framed. The plan-conditioned risk gating with a handoff back to the base follower is a practical design.

Methodological Rigor

Design: The modular architecture is reasonable. Supervision is clean (privileged annotations used only in training), and train and eval episodes are disjoint.

Ablations: They are extensive and informative. They cover single-frame input, no beliefs, no occlusion-time propagation, no waypoint conditioning, no recurrent memory, reactive triggering, and always-on control. The paper acknowledges that some ablations are confounded (the single-frame variant also disables propagation, and removing both beliefs does not isolate either one).

Baselines: The comparison is thin. OA-VAT is the only strong occlusion-aware baseline. The VLA baselines are used off-the-shelf rather than retrained on OccEVT-Expert. The 121.4% headline is relative to OA-VAT, and the improvement over the base follower (1% → 77.5%) is partly an artifact of training on in-distribution expert data for this specific benchmark.

Evaluation fairness: The benchmark is authors-designed, and Risk episodes are defined by occlusion under the reference follower (OmTrackVLA). That definition favors a method built on that follower and is arguably adversarial to it. Baselines were not trained on the 20k demonstrations.

Statistics: There are only 400 episodes (about 200 per split), with no confidence intervals or seeds. Duration metrics are confounded by early collision termination, which the authors admit.

Safety: On Risk episodes CFR (90.5%) is below OmTrackVLA's (97.5%). On Safe episodes VP-EVT is no better than the base follower. Real-world validation is a single qualitative demonstration on a Go2 with an off-board GPU.

Sim-to-real: The sim-to-real gap is not quantified.

Potential Impact

Direct audience: The audience is the robot person-following and EVT community, a fairly narrow subfield. The benchmark's attribution metrics (POL versus OOF, LTL) are useful and may be adopted by others studying occlusion-aware following.

Use cases: Applications include service robots, logistics, and assistive following.

Reuse: The "monitor, predict, intervene, hand back" structure over a frozen follower is reusable. It could extend to other embodied tasks that require active perception.

Dependencies: Impact depends on whether OccEVT-Bench and the dataset are actually released, and on Habitat-GS adoption. The paper does not state a release.

Timeliness & Relevance

The paper fits the current wave of VLA-based tracking (TrackVLA, TrackVLA++, OmTrackVLA), which has focused on perception, memory, and recovery rather than active occlusion avoidance. Persistent occlusion in crowds is a real bottleneck, and the evaluation gap the paper identifies (aggregate success hiding persistent target loss) is legitimate. The topic is relevant but incremental within robotics.

Strengths

Clear problem formulation that separates physical occlusion from out-of-FOV loss.

Large effect sizes and a coherent ablation story supporting predictive, selective intervention.

Modular and plug-in over an existing follower, with no regression on Safe episodes.

Benchmark and dataset contributions with thoughtful metric design.

Honest limitations discussion.

Limitations

The benchmark is self-designed and Risk episodes are constructed around the base follower's failure, so the headline gains may be inflated.

The baselines are not retrained on the expert data, which confounds the method and data contributions.

There are no statistical uncertainty estimates and a small test set.

Real-world evidence is anecdotal.

Collision safety is lower than the base follower's on Risk episodes.

The method is complex, with many learned modules and staged training. Reproducibility is moderate, since some hyperparameters and the release status are unclear.

Generalization to denser crowds, multiple occluders, and non-pedestrian occluders is untested.

The approach relies on privileged expert demonstrations generated in simulation.

Overall

This is a solid, well-organized robotics paper with a clear conceptual framing, a useful benchmark, and strong ablations. Its influence is likely confined to the EVT and person-following subcommunity, and the evaluation advantages are somewhat self-constructed. Expected impact is moderate.

Rating:5.5/ 10

Significance 5Rigor 5.5Novelty 5.5Clarity 7

Generated Oct 16, 2026

Comparison History (0)

No comparisons yet.

Scorecard· 16/16

5.5/10 impact