Pith. sign in

REVIEW 3 major objections 5 minor 24 references

End-to-end driving systems that keep actor hypotheses alive through full occlusion recover hidden agents and cut open-loop planning error.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-13 05:07 UTC pith:6BOPKHNH

load-bearing objection Solid engineering paper that makes object permanence a first-class requirement in E2E driving, with a usable benchmark; the 0.249 unobs mAP is softer than the abstract implies because of adaptive matching, but the planning gains still hold. the 3 major comments →

arxiv 2607.09138 v1 pith:6BOPKHNH submitted 2026-07-10 cs.RO

BeyondSight: Object Permanence for End-to-End Autonomous Driving

Sandro Papais , Letian Wang , Mudit Jain , Behnaz Rezaei , Steven L. Waslander This is my paper
classification cs.RO
keywords object permanenceend-to-end autonomous drivingpartial observabilityocclusionsparse queriesspatiotemporal reasoningnuScenes-Permanence
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Most end-to-end autonomous driving stacks treat an actor as real only while sensors can still see it. When a pedestrian or vehicle vanishes behind another car, its hypothesis is dropped, so prediction and planning never reason about it. BeyondSight decouples existence from instantaneous visibility: it propagates actor queries through time without needing image evidence, then fuses them with any new observations so the scene representation always contains both visible and temporarily hidden agents. The authors also build nuScenes-Permanence, which supplies labels and metrics for fully unobservable actors so this skill can be trained and scored fairly. On that protocol the method raises unobservable detection from zero to 0.249 mAP and lowers average planning L2 from 0.61 to 0.54, while standard visible-actor metrics stay competitive. The result is that object permanence is not merely a developmental concept; it is a concrete modeling requirement for safer closed-loop behavior under occlusion.

Core claim

Decoupling actor existence from instantaneous observability—by maintaining temporally propagated, observation-conditioned actor hypotheses that survive full occlusion—improves joint perception, prediction, and planning. On nuScenes-Permanence, BeyondSight lifts unobservable-actor mAP from 0 to 0.249 and reduces open-loop planning L2avg from 0.61 to 0.54 without harming standard visible-actor scores.

What carries the argument

A three-stage actor belief update: a Temporal Prior Decoder that propagates previous queries without image features, an Observation Decoder that conditions on current BEV features, and a Posterior Fusion Decoder that retains the higher-confidence hypothesis—paired with permanence-aware supervision over both observable and unobservable ground truth.

Load-bearing premise

The offline reconstruction of occluded trajectories—by interpolation and short-horizon forecasting—is accurate enough to serve as ground truth for both training and evaluation.

What would settle it

Train and evaluate the identical architecture on a multi-observer or privileged-sensor dataset that records true 3D states of fully occluded actors without reconstruction; if mAPunobs collapses and the planning L2 gains vanish, the permanence claim fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. BeyondSight proposes a permanence-aware end-to-end driving stack that decouples actor existence from instantaneous observability by propagating sparse actor queries through a Temporal Prior Decoder, updating them with an Observation Decoder, and reconciling them in a Posterior Fusion Decoder (Eqs. 3–8). The authors also introduce nuScenes-Permanence, which retains zero-point boxes and completes unobservable trajectory intervals offline, together with an observability-conditioned evaluation protocol (mAPobs/unobs/all). On standard nuScenes, BeyondSight improves open-loop planning from 0.61 to 0.54 L2avg and slightly improves detection/tracking over SparseDrive and related baselines (Tables 1–2). On nuScenes-Permanence, unobservable detection rises from 0 to 0.249 mAPunobs (Table 3), with ablations attributing gains to temporal prior, fusion, and unobservable supervision (Table 4).

Significance. Object permanence under full occlusion is a genuine and under-addressed requirement for end-to-end autonomous driving; most current stacks and benchmarks equate existence with sensor support. The paper’s contribution is both architectural (explicit prior/observation/fusion query stages) and infrastructural (nuScenes-Permanence annotations and observability-conditioned metrics). If the permanence gains hold under tighter evaluation, the work would establish a useful modeling principle and a reusable benchmark extension. Strengths include clear ablations isolating temporal prior, fusion, and unobservable supervision; a high-occlusion planning subset; occlusion-duration stratification; and competitive standard-protocol metrics showing that persistence does not collapse official mAP/NDS. The planning L2 and collision improvements without changing the planner itself are practically relevant.

major comments (3)
  1. [Table 3; Supp. D Eqs. (17)–(18); Supp. C Table 10] The headline unobservable detection claim (abstract; Table 3: mAPunobs 0 → 0.249) depends on the adaptive matching threshold in Supp. D, Eqs. (17)–(18), whose coefficients are fit by 90th-percentile quantile regression on the same offline completion pipeline used to create unobservable labels (Supp. C, Table 10). Reconstruction L2 already reaches mean 1.11 m / P90 3.22 m for 4–6 s gaps. Because the tolerance is deliberately enlarged with velocity/acceleration terms, a non-trivial fraction of counted true positives may be loose matches to noisy extrapolated boxes rather than accurate persistent hypotheses. Please report mAPunobs under fixed nuScenes thresholds (or with tolerance ablated to zero extra slack) and under zero-point-only / interpolation-only subsets, so the permanence gain can be separated from matching slack.
  2. [Supp. Table 8; §6.2 Object Permanence Performance] Supp. Table 8 reports very high false-discovery rates for unobservable predictions (FDRunobs 0.85 at 0–2 s, rising to 0.98 at 4–6 s) while TPRunobs falls from 0.62 to 0.35. This is load-bearing for the claim that BeyondSight “maintains persistent actor hypotheses”: high FDR implies many maintained unobservable queries do not correspond to valid unobservable actors. The main text does not discuss this precision failure mode. Please surface FDR/TPR (or precision–recall) for unobservable actors in the main results, analyze pruning/confidence calibration of stale hypotheses, and clarify how high FDR interacts with the reported mAPunobs and planning gains.
  3. [§5.1; Supp. C; Table 4] Unobservable supervision and evaluation rest on offline trajectory completion (constant-turn-rate/acceleration interpolation plus a SparseDrive-derived short-horizon forecaster for terminal gaps; §5.1, Supp. C). The paper correctly notes that abrupt hidden maneuvers are out of scope and that error grows with horizon, but training and adaptive evaluation both use this generated GT. Mild circularity risk remains because the completion model is architecturally related to the SparseDrive family. Please quantify how much of the Table 4 gain survives when (i) only zero-point retained boxes are used for unobs supervision (no interp/extrap), and (ii) evaluation ignores extrapolated terminal states. Without that, the 0.249 figure is hard to interpret as pure permanence rather than fitting the completion prior.
minor comments (5)
  1. [Fig. 1; Fig. 3] Figure 1 and Figure 3 are persuasive but would benefit from explicit timestamps and occlusion duration annotations so readers can judge “prolonged” occlusion quantitatively.
  2. [§3 Eq. (2)] Eq. (2) formalizes permanence as existence of a representation ri_t, but does not constrain localization quality or confidence; a short remark linking this definition to the observability head and pruning would tighten the formalism.
  3. [Table 2] Table 2 reports minADE identical to SparseDrive (0.610) while EPA improves slightly; a one-sentence interpretation of why permanence helps EPA more than ADE would help.
  4. [Supp. B Latency Analysis] Latency is reported only in the supplement (6.1 → 6.0 FPS). A brief main-text note that permanence adds <5% parameters and negligible runtime would strengthen the practicality claim.
  5. [Abstract; Table 4] Minor wording: abstract and intro say “increasing detection performance for unobservable actors from 0 to 0.249 mAP” without immediately noting that SparseDrive’s zero is protocol-induced under standard filtering; Table 4’s SparseDrive-on-Permanence baseline (0.021) should be cited alongside the abstract claim.

Circularity Check

1 steps flagged

Mild risk that adaptive matching tolerances fitted to the same offline completion pipeline inflate mAPunobs; core gains still hold vs. baselines under identical protocol and on planning metrics.

specific steps
  1. fitted input called prediction [Supp. D Adaptive Matching Threshold (Eqs. 17–18); used for Table 3 mAPunobs]
    "To avoid conflating model error with extrapolation drift, we augment the standard nuScenes matching threshold with an adaptive tolerance: d_match(v, a, t, ℓ) = d_nuScenes(ℓ) + d̂_ℓ(v, a, t), ... The coefficients (α_ℓ, β_ℓ, γ_ℓ) are estimated using non-negative 90th-percentile quantile regression on approximately 1.5×10^6 extrapolated box–horizon pairs generated by masking observable nuScenes annotations."

    The key reported figure mAPunobs = 0.249 is computed with a matching radius whose extra slack is fitted directly to reconstruction errors of the identical offline trajectory-completion process that creates the unobservable labels. Longer-horizon (noisier) labels automatically receive larger tolerance, so a non-trivial fraction of “true positives” can be loose matches enabled by the fitted slack rather than precise persistent hypotheses. Baselines still score 0, so the result is not fully forced, but the metric is partially self-calibrated to the label generator.

full rationale

BeyondSight is an empirical end-to-end driving paper whose central claims are measured improvements (mAPunobs 0 o0.249, L2avg 0.61 o0.54) on held-out nuScenes validation scenes and a self-constructed permanence extension. There is no mathematical derivation chain, uniqueness theorem, or self-definitional loop that forces the result by construction. The sole mild circularity concern is that both the unobservable ground-truth labels (Supp. C: interpolation + SparseDrive-adapted forecaster) and the adaptive matching radius used for mAPunobs (Supp. D Eqs. 17–18, 90th-percentile quantile regression on the same reconstruction errors) originate from the identical offline pipeline. This can loosen the criterion in proportion to label noise, especially for longer occlusions. However, SparseDrive scores exactly 0 under the same protocol, ablations show progressive gains only when permanence components are added, and planning/collision metrics (independent of detection matching) also improve on a high-occlusion subset. The annotation model is never used at inference. Thus the circularity is limited to evaluation calibration rather than the model’s learned behavior, warranting only a low score of 2.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 2 invented entities

The central empirical claim rests on a small set of domain conventions (LiDAR-point observability, SparseDrive backbone) plus a handful of hand-chosen training hyperparameters and the offline trajectory-completion procedure that defines the new benchmark. No new physical entities are postulated; the invented modules are architectural components whose value is measured by ablation.

free parameters (3)
  • λ_obs (observability loss weight) = 1.0
    Set to 1.0 by hand; controls how strongly the model is forced to predict the binary observability state.
  • Stage-1 / Stage-2 learning rates and backbone multipliers = 4e-4 / 3e-4
    4e-4 / 3e-4 with 0.5 / 0.1 backbone multipliers; standard but still free choices that affect final numbers.
  • Adaptive matching coefficients (α_ℓ, β_ℓ, γ_ℓ) = class-specific (Table 13)
    Fitted by non-negative 90th-percentile quantile regression on ~1.5 M masked boxes; directly used in mAP_unobs evaluation.
axioms (3)
  • domain assumption An actor is unobservable exactly when it has zero associated LiDAR/radar points (Eq. 1).
    Standard nuScenes convention adopted without further sensor modeling; defines both the problem and the benchmark split.
  • domain assumption Actor motion during unobservable intervals is approximately continuous and can be recovered by constant-turn-rate interpolation or short-horizon forecasting.
    Stated in Sec. 5.1; abrupt hidden maneuvers are explicitly declared out of scope.
  • domain assumption SparseDrive’s sparse-query detection, motion and planning heads remain a valid backbone once permanence modules are added.
    BeyondSight is defined as an extension of SparseDrive; all comparisons inherit that architecture.
invented entities (2)
  • Temporal Prior Decoder + Posterior Fusion Decoder no independent evidence
    purpose: Propagate actor queries without image features and reconcile them with observation-conditioned queries so that unobservable actors remain in the scene representation.
    New architectural modules; their contribution is measured by ablation (Table 4) rather than by independent physical evidence.
  • nuScenes-Permanence annotation set and observability-conditioned metrics no independent evidence
    purpose: Supply supervision and evaluation for fully unobservable actors that standard nuScenes discards.
    Constructed offline by retaining zero-point boxes and completing trajectories; quality is validated by hold-out reconstruction but remains synthetic.

pith-pipeline@v1.1.0-grok45 · 20370 in / 2922 out tokens · 29365 ms · 2026-07-13T05:07:46.042581+00:00 · methodology

0 comments
read the original abstract

Autonomous driving operates in partially observable environments where actors may become fully occluded by other vehicles or infrastructure. Most end-to-end driving systems implicitly couple actor existence to instantaneous observations, causing actor hypotheses to degrade or disappear during prolonged occlusion and removing potentially critical agents from downstream prediction and planning. We introduce BeyondSight, a permanence-aware end-to-end driving framework that decouples actor existence from observability by maintaining persistent actor hypotheses over time. BeyondSight propagates actor queries temporally and updates them with observation-conditioned evidence, enabling joint perception, prediction, and planning to reason about actors even when they are temporarily unobservable. To enable principled training and evaluation of persistence-aware models, we further introduce nuScenes-Permanence, an extension of nuScenes that provides supervision and observability-conditioned evaluation for unobservable actors. Experiments show that BeyondSight substantially improves reasoning under occlusion, increasing detection performance for unobservable actors from 0 to 0.249 mAP while reducing planning error from 0.61 to 0.54 L2avg. These results highlight object permanence as an important modeling principle for robust end-to-end autonomous driving.

Figures

Figures reproduced from arXiv: 2607.09138 by Behnaz Rezaei, Letian Wang, Mudit Jain, Sandro Papais, Steven L. Waslander.

Figure 1
Figure 1. Figure 1: Object permanence in prolonged occlusion. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: BeyondSight overview. Actor belief update consists of three stages: (1) a Tem￾poral Prior Decoder propagates actor hypotheses from the previous frame, (2) an Ob￾servation Decoder updates hypotheses using current image features, and (3) a Posterior Fusion Decoder merges and refines propagated and observation-conditioned hypothe￾ses. 4.1 BeyondSight Model To satisfy the actor persistence constraint defined i… view at source ↗
Figure 3
Figure 3. Figure 3: Planning under full occlusion. Left: SparseDrive drops the occluded actor once visual evidence disappears, causing the ego plan to intersect the actor’s ground-truth future trajectory. Right: BeyondSight maintains a persistent hypothesis for the hidden actor and produces a safer plan. The occluder, hidden actor, actor future, and ego plan are highlighted for clarity. To quantify this behavior beyond a sing… view at source ↗
Figure 4
Figure 4. Figure 4: Distribution of full-occlusion durations (left) and ego distance (right) in nuScenes. but fully unobservable. Since such intervals are not consistently retained under the standard nuScenes filtering rules, we construct additional annotations offline using a trajectory completion procedure. For each actor track, we identify contiguous intervals of missing observability and estimate the latent actor state du… view at source ↗
Figure 5
Figure 5. Figure 5: Examples of generated annotations including full unobservability for a single scene. Boxes are colored according to their observability status and dataset generation method. Observability-Conditioned Evaluation. At each timestamp the ground￾truth actor set is partitioned into observable and unobservable subsets G_{\mathrm {obs}}, \qquad G_{\mathrm {unobs}}. Evaluation is performed by selecting a target set… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

24 extracted references · 2 linked inside Pith

  1. [1]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621–11631 (2020)

    2020

  2. [2]

    In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision

    Cheong, B., Wang, L., Papais, S., Waslander, S.L.: Scatr: Mitigating new instance suppression in lidar-based tracking-by-attention via second chance assignment and track query dropout. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3330–3339 (2026)

    2026

  3. [3]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Gu, J., Hu, C., Zhang, T., Chen, X., Wang, Y., Wang, Y., Zhao, H.: Vip3d: End- to-end visual trajectory prediction via 3d agent queries. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5496– 5506 (2023)

    2023

  4. [4]

    Hu, S., Chen, L., Wu, P., Li, H., Yan, J., Tao, D.: St-p3: End-to-end vision-based autonomousdrivingviaspatial-temporalfeaturelearning.In:EuropeanConference on Computer Vision. pp. 533–549. Springer (2022)

    2022

  5. [5]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Hu, Y., Yang, J., Chen, L., Li, K., Sima, C., Zhu, X., Chai, S., Du, S., Lin, T., Wang, W., et al.: Planning-oriented autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17853– 17862 (2023)

    2023

  6. [6]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023)

    Jia, X., Gao, Y., Chen, L., Yan, J., Liu, P.L., Li, H.: DriveAdapter: Breaking the coupling barrier of perception and planning in end-to-end autonomous driving. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023)

    2023

  7. [7]

    arXiv preprint arXiv:2503.07656 (2025)

    Jia, X., You, J., Zhang, Z., Yan, J.: Drivetransformer: Unified transformer for scalable end-to-end autonomous driving. arXiv preprint arXiv:2503.07656 (2025)

    Pith/arXiv arXiv 2025

  8. [8]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Jiang, B., Chen, S., Xu, Q., Liao, B., Chen, J., Zhou, H., Zhang, Q., Liu, W., Huang,C.,Wang,X.:Vad:Vectorizedscenerepresentationforefficientautonomous driving. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8340–8350 (2023)

    2023

  9. [9]

    arXiv preprint arXiv:2409.18341 (2024)

    Li, P., Cui, D.: Navigation-guided sparse scene representation for end-to-end au- tonomous driving. arXiv preprint arXiv:2409.18341 (2024)

    Pith/arXiv arXiv 2024

  10. [10]

    In: European conference on computer vision

    Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird’s-eye-view representation from multi-camera images via spatiotem- poral transformers. In: European conference on computer vision. pp. 1–18. Springer (2022)

    2022

  11. [11]

    In: 2024 IEEE International Conference on Robotics and Automation (ICRA)

    Papais, S., Ren, R., Waslander, S.: Swtrack: Multiple hypothesis sliding window 3d multi-object tracking. In: 2024 IEEE International Conference on Robotics and Automation (ICRA). pp. 4939–4945. IEEE (2024)

    2024

  12. [12]

    In: Proceedings of the IEEE/CVF 16 S

    Papais, S., Wang, L., Cheong, B., Waslander, S.L.: Foresight: Multi-view streaming joint object detection and trajectory forecasting. In: Proceedings of the IEEE/CVF 16 S. Papais et al. International Conference on Computer Vision (ICCV). pp. 25474–25484 (October 2025)

    2025

  13. [13]

    In: European Conference on Computer Vision (ECCV)

    Philion, J., Fidler, S.: Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In: European Conference on Computer Vision (ECCV). pp. 194–210 (2020).https://doi.org/10.1007/978-3-030-58568-6_12

    doi:10.1007/978-3-030-58568-6_12 2020

  14. [14]

    In: European Conference on Computer Vision

    Shamsian, A., Kleinfeld, O., Globerson, A., Chechik, G.: Learning object per- manence from video. In: European Conference on Computer Vision. pp. 35–50. Springer (2020)

    2020

  15. [15]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Shao, H., Wang, L., Chen, R., Waslander, S.L., Li, H., Liu, Y.: Reasonnet: End-to- end driving with temporal and global reasoning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13723–13733 (2023)

    2023

  16. [16]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Song, Z., Jia, C., Liu, L., Pan, H., Zhang, Y., Wang, J., Zhang, X., Xu, S., Yang, L., Luo, Y.: Don’t shake the wheel: Momentum-aware planning in end-to-end au- tonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 22432–22441 (June 2025)

    2025

  17. [17]

    In: 2025 IEEE International Conference on Robotics and Automation (ICRA)

    Sun, W., Lin, X., Shi, Y., Zhang, C., Wu, H., Zheng, S.: Sparsedrive: End-to-end autonomous driving via sparse scene representation. In: 2025 IEEE International Conference on Robotics and Automation (ICRA). pp. 8795–8801. IEEE (2025)

    2025

  18. [18]

    In: International Conference on Machine Learning

    Tokmakov, P., Jabri, A., Li, J., Gaidon, A.: Object permanence emerges in a random walk along memory. In: International Conference on Machine Learning. pp. 21506–21519. PMLR (2022)

    2022

  19. [19]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Tokmakov, P., Li, J., Burgard, W., Gaidon, A.: Learning to track with object per- manence. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10860–10869 (2021)

    2021

  20. [20]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Van Hoorick, B., Tokmakov, P., Stent, S., Li, J., Vondrick, C.: Tracking through containers and occluders in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13802–13812 (2023)

    2023

  21. [21]

    Foundations and Trends®in Robotics13(1-2), 1–269 (2026)

    Wang, L., Lavoie, M.A., Papais, S., Nisar, B., Chen, Y., Ding, W., Ivanovic, B., Shao, H., Abuduweili, A., Cook, E., et al.: Trends in motion prediction toward deployable and generalizable autonomy: A revisit and perspectives. Foundations and Trends®in Robotics13(1-2), 1–269 (2026)

    2026

  22. [22]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Wang, S., Liu, Y., Wang, T., Li, Y., Zhang, X.: Exploring object-centric tempo- ral modeling for efficient multi-view 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3621–3631 (2023)

    2023

  23. [23]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Xue, Z., Guo, M., Fan, H., Zhang, S., Zhang, Z.: Corrbev: Multi-view 3d object detection by correlation learning with multi-modal prototypes. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 27413–27423 (2025)

    2025

  24. [24]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Zhang, B., Song, N., Jin, X., Zhang, L.: Bridging past and future: End-to-end autonomous driving with historical prediction and planning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 6854–6863 (June 2025) BeyondSight: Object Permanence for End-to-End Autonomous Driving 17 Supplementary Material A Additi...

    2025