Pith. sign in

REVIEW 2 major objections 5 minor 57 references

DeLL grows dual knowledge spaces and applies front-door causal adjustment so end-to-end driving models can learn new scenarios without erasing old skills.

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

2026-07-14 21:20 UTC pith:5TOXG5LN

load-bearing objection Solid systems paper that actually delivers lifelong gains in closed-loop CARLA; the front-door story is the softest part but the empirical package still holds. the 2 major comments →

arxiv 2603.14354 v3 pith:5TOXG5LN submitted 2026-03-15 cs.LG cs.AIcs.RO

Deconfounded Lifelong Learning for Autonomous Driving via Dynamic Knowledge Spaces

Jiayuan Du , Yuebing Song , Yiming Zhao , Xianghui Pan , Jiawei Lian , Yuchu Lu , Liuyi Wang , Chengju Liu
show 1 more author
Qijun Chen
This is my paper
classification cs.LG cs.AIcs.RO
keywords end-to-end autonomous drivinglifelong learningDirichlet process mixture modelfront-door adjustmentcatastrophic forgettingcausal inferenceCARLABench2Drive
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.

End-to-end driving systems trained by imitation tend to overwrite earlier skills when new ones arrive and to latch onto spurious links between what sensors show and what the car should do. This paper claims that a Dirichlet process mixture model can keep two expanding knowledge libraries—one of latent scene features, one of expert trajectories—so new patterns create new anchors without destroying old ones. Those anchors then serve as mediators in a front-door adjustment that strips out unobserved confounders such as sensor noise or weather shifts before planning. An evolutionary decoder scores and refines the growing set of trajectory anchors in parallel rather than step by step. On a new lifelong protocol built from Bench2Drive tasks inside CARLA, the method raises average driving score, lowers forgetting, and improves both forward and backward transfer, and it also leads under ordinary full-data training.

Core claim

A single framework that couples non-parametric dual knowledge spaces with attention-based front-door adjustment lets an end-to-end driving stack expand its repertoire of abilities across sequential tasks while retaining prior skills and reducing decisions driven by unobserved confounders.

What carries the argument

DPMM-based dual dynamic knowledge spaces (feature anchors and trajectory anchors) that act as the mediators M in an attention approximation of front-door adjustment, followed by an evolutionary non-autoregressive trajectory decoder.

Load-bearing premise

The growing cluster centers are assumed to be pure enough mediators that routing through them fully blocks the effect of unobserved confounders between perception and action.

What would settle it

Keep the expanding DPMM anchors but turn off the front-door attention modules; if forgetting ratio and transfer scores stay essentially the same on the lifelong Bench2Drive sequence, the deconfounding claim is unnecessary. Alternatively, inject a controlled CARLA confounder (systematic camera bias or weather) that leaves cluster assignments unchanged; if performance then collapses, the front-door step has failed to remove residual confounding.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • New driving abilities can be added without replay buffers or fixed task boundaries while emergency-braking and traffic-sign skills are retained.
  • Average driving score, success rate and multi-ability mean improve under both lifelong and full-data training on the proposed Bench2Drive protocol.
  • Spurious correlations induced by sensor noise or environmental change are reduced by treating knowledge anchors as front-door mediators.
  • Trajectory planning can be performed non-autoregressively by scoring and refining a dynamically growing set of kinematic anchors.
  • The introduced lifelong metrics (forgetting ratio, process forgetting, forward and backward transfer) become a reusable evaluation suite for other closed-loop E2E-AD methods.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same dual-space-plus-front-door pattern could be tried in other continual control settings (manipulation, aerial navigation) where skills arrive sequentially and confounders are unobserved.
  • If the birth/merge heuristics remain stable on real logs, fleets could maintain a shared growing knowledge library without shipping private replay data to a central server.
  • Attention-as-front-door may offer a lightweight way to inject causal structure into transformer planners when a full causal graph is unavailable.
  • A diagnostic for residual confounding would be any distribution shift that leaves DPMM assignments intact yet still degrades closed-loop performance after the front-door modules are applied.

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

2 major / 5 minor

Summary. The paper proposes DeLL, a lifelong learning framework for end-to-end autonomous driving that couples Dirichlet process mixture models (DPMM) with front-door adjustment. DPMM maintains two expanding knowledge spaces—feature anchors (FKS) and trajectory anchors (TKS)—that grow without a fixed cluster count and serve as mediators in cascaded attention modules (FFEM/TFEM) intended to approximate Pearl’s front-door formula. An evolutionary trajectory decoder then performs non-autoregressive, Top-K planning over the trajectory anchors. The authors introduce a sequential Bench2Drive protocol (five ability categories ordered by decreasing data volume) together with forgetting, process-forgetting, forward- and backward-transfer metrics. Closed-loop CARLA results show improved average driving score (74.69 %), reduced forgetting (FR 33.97 %), and higher backward transfer (79.63 %) relative to a TF++ baseline, with further gains under full-data training (DS 86.86 %) and competitive comparisons against adapted ER and PackNet.

Significance. If the empirical gains hold, the work supplies the first systematic lifelong-learning benchmark and architecture for closed-loop E2E-AD in CARLA, together with a practical non-parametric memory mechanism that avoids both fixed task boundaries and large replay buffers. The dual knowledge spaces, evolutionary decoder, and new evaluation protocol are concrete engineering contributions that other groups can reuse. The causal front-door story is more aspirational than rigorously verified, yet the ablation and reverse-order experiments still demonstrate that the DPMM-based expansion alone yields substantial retention and transfer benefits. Code release further raises the paper’s utility for the community.

major comments (2)
  1. Section 3.3 and Eq. (2): the central causal claim treats DPMM anchors A_feat and A_traj as mediators M that satisfy the front-door criterion (X o M o Y with no unblocked back-door from M to Y). No diagnostic is provided that residual confounders (sensor noise, weather, map layout, unobserved agent intents) are blocked once features are projected onto these clusters. Table 3 shows that removing FFEM/TFEM still leaves large gains over TF++, consistent with ordinary non-parametric memory expansion rather than genuine deconfounding. The manuscript should either (i) supply a concrete test of residual confounding or (ii) soften the “deconfounded” language to “knowledge-mediated feature enhancement.”
  2. Section 4.1 and the lifelong protocol: the five Bench2Drive ability categories ordered by decreasing data volume are presented as a realistic lifelong sequence, yet no sensitivity analysis to alternative orderings (beyond the reverse sequence in the supplement) or to continuous rather than discrete ability boundaries is reported. Because the forgetting and transfer metrics are defined with respect to this particular ordering, the quantitative claims in Table 1 remain protocol-dependent until additional orderings or a continuous stream are examined.
minor comments (5)
  1. Notation inconsistency: F_fused is written both as R^{11 imes256} and later as R^{11 imes256} with a trailing prime; a single consistent symbol for the enhanced features would improve readability.
  2. Figure 4 caption and main text: the t-SNE visualization of feature clusters is informative, yet the color-to-ability mapping is never listed explicitly; a short legend would help readers interpret the claimed multi-cluster structure of each ability.
  3. Eq. (5) and surrounding text: the temperature au and Top-K hyper-parameters are free but never ablated; a one-sentence sensitivity note would strengthen the decoder claims.
  4. Related-work section: recent continual-learning methods for trajectory prediction (e.g., H2C, LiLoDriver) are cited only in the bibliography; a short comparative paragraph would better situate the contribution.
  5. Typographical: “multi-sacale” (p. 4), “mis-sion” (p. 6), and occasional missing spaces after periods appear throughout; a careful proof-reading pass is needed.

Circularity Check

0 steps flagged

No significant circularity: empirical systems paper whose lifelong-learning gains are measured on held-out CARLA routes, not algebraic rearrangements of fitted inputs.

full rationale

DeLL is an architectural and empirical contribution. The dual DPMM knowledge spaces, attention-based front-door modules (Eqs. 3–4 approximating Pearl’s external formula Eq. 2), and evolutionary trajectory decoder (Eq. 5) are design choices whose effects are evaluated by closed-loop driving score, success rate, FR, PFR, FT and BT on sequential Bench2Drive tasks (Tables 1–3, reverse-order Table 5, full-data Table 2). These metrics are computed from simulator outcomes after training; they are not defined in terms of the DPMM cluster centers or the attention weights, nor are they recovered by construction from a fitted parameter. Front-door validity of the anchors is an untested modeling assumption (a correctness risk), not a circular reduction. Citations to Pearl, memoVB, Transfuser++ and standard lifelong-learning baselines are external; no load-bearing uniqueness theorem or ansatz is imported from overlapping authors. Ablations further isolate module contributions rather than tautologically restate the full model. Consequently the derivation chain contains no self-definitional step, no fitted-input-called-prediction, and no self-citation circularity.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 3 invented entities

The central empirical claim rests on standard Bayesian non-parametrics and Pearl’s front-door formula plus several modeling choices that are not independently validated outside the CARLA experiments. Free parameters are the usual training hyper-parameters plus DPMM concentration and temperature; invented entities are the two knowledge spaces and the evolutionary decoder.

free parameters (4)
  • DPMM concentration α
    Controls birth of new clusters; value not reported but directly affects how many knowledge anchors appear and therefore the capacity of the lifelong memory.
  • temperature τ in Softmax for trajectory probabilities
    Scales the anchor-selection distribution (Eq. 5 and 8); chosen by hand and influences both training loss and Top-K routing.
  • number of learnable queries (11) and Top-K
    Architectural constants that determine the size of the fused representation and the final trajectory set; fixed without ablation on sensitivity.
  • learning-rate schedule (3e-4 → 3e-5) and epoch counts
    Standard optimizers settings that affect final numbers reported in Tables 1–3.
axioms (4)
  • standard math Dirichlet-process mixture model generative process (Eq. 1) with Normal-Wishart base and diagonal-covariance Gaussians yields valid, expandable clusters of driving features and trajectories.
    Invoked throughout Section 3.2; standard Bayesian non-parametrics but the diagonal-covariance restriction is a computational approximation.
  • standard math Front-door adjustment formula (Eq. 2) identifies P(Y|do(X)) when M intercepts all directed paths and no unblocked back-door exists from M to Y.
    Taken from Pearl and applied in Section 3.3; the paper assumes the DPMM anchors satisfy the graphical criteria.
  • domain assumption Unobserved confounders (sensor noise, environmental shifts) act only through the back-door X ← U → Y and are blocked once the anchors are used as mediators.
    Stated in the introduction and Section 3.3; not independently verified.
  • ad hoc to paper The five Bench2Drive ability categories ordered by decreasing data volume constitute a realistic lifelong task sequence for measuring forgetting and transfer.
    Defined in Section 4.1; the ordering and metrics (FR, PFR, FT, BT) are new constructions of the authors.
invented entities (3)
  • Feature Knowledge Space (FKS) and Trajectory Knowledge Space (TKS) no independent evidence
    purpose: Dynamically store and expand latent feature patterns and explicit trajectory prototypes as knowledge anchors for lifelong retention and causal mediation.
    Core architectural novelty of Section 3.2; no external independent evidence beyond the CARLA ablations.
  • Evolutionary trajectory decoder with dual-branch Top-K routing no independent evidence
    purpose: Map growing trajectory anchors into parallel non-autoregressive plans.
    Introduced in Section 3.4 to match the dynamic knowledge base; performance gains shown only inside the same experiments.
  • Cascaded FFEM/TFEM attention modules implementing front-door adjustment no independent evidence
    purpose: Re-express current features as mixtures of historical anchors to remove confounding.
    Section 3.3; the attention approximation of the front-door sum is paper-specific.

reviewed 2026-07-14 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Deconfounded Lifelong Learning for Autonomous Driving via Dynamic Knowledge Spaces." pith.science (2026). https://pith.science/paper/5TOXG5LN

@misc{pith2026260314354,
  author       = {Pith},
  title        = {Pith review of: Deconfounded Lifelong Learning for Autonomous Driving via Dynamic Knowledge Spaces},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5TOXG5LN}},
  note         = {Machine review of arXiv:2603.14354}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

End-to-End autonomous driving (E2E-AD) systems face challenges in lifelong learning, including catastrophic forgetting, difficulty in knowledge transfer across diverse scenarios, and spurious correlations between unobservable confounders and true driving intents. To address these issues, we propose DeLL, a Deconfounded Lifelong Learning framework that integrates a Dirichlet process mixture model (DPMM) with the front-door adjustment mechanism from causal inference. The DPMM is employed to construct two dynamic knowledge spaces: a trajectory knowledge space for clustering explicit driving behaviors and an implicit feature knowledge space for discovering latent driving abilities. Leveraging the non-parametric Bayesian nature of DPMM, our framework enables adaptive expansion and incremental updating of knowledge without predefining the number of clusters, thereby mitigating catastrophic forgetting. Meanwhile, the front-door adjustment mechanism utilizes the DPMM-derived knowledge as mediators to deconfound spurious correlations, such as those induced by sensor noise or environmental changes, and enhances the causal expressiveness of the learned representations. Additionally, we introduce an evolutionary trajectory decoder that enables non-autoregressive planning. To evaluate the lifelong learning performance of E2E-AD, we propose new evaluation protocols and metrics based on Bench2Drive. Extensive evaluations in the closed-loop CARLA simulator demonstrate that our framework significantly improves adaptability to new driving scenarios and overall driving performance, while effectively retaining previously acquired knowledge. Code: https://github.com/Mooncakebro/DeLL

Figures

Figures reproduced from arXiv: 2603.14354 by Chengju Liu, Jiawei Lian, Jiayuan Du, Liuyi Wang, Qijun Chen, Xianghui Pan, Yiming Zhao, Yuchu Lu, Yuebing Song.

Figure 1
Figure 1. Figure 1: Driving score, success rate and multi-ability success rate during the lifelong learning process. Our method not only demonstrates superior overall driving perfor￾mance but also substantially mitigates catastrophic forgetting. shown in [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview architecture of our proposed method. We also introduce auxiliary learning tasks (BEV semantic segmentation and detection) to enrich the BEV representations with clear geometric and semantic boundaries. These geometrically constrained BEV features are then transformed and concatenated with the ego-vehicle’s current velocity feature and target point feature. Finally, a transformer decoder equipped w… view at source ↗
Figure 3
Figure 3. Figure 3: Lifelong performance on CARLA benchmark DEV10 [21]. As shown in [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: presents the clustering results of dynamic knowledge spaces during the learning process, including numbers, IDs and visualizations. In particular, we sample 50 data points per cluster from the feature knowledge space and project them into 2D using t-SNE [27] for visualization, with colors indicating their associated driving ability. Notably, some driving capabilities encompass multiple clusters. While most… 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

57 extracted references · 14 linked inside Pith

  1. [1]

    In: ECCV

    Aljundi, R., Babiloni, F., Elhoseiny, M., Rohrbach, M., Tuytelaars, T.: Memory aware synapses: Learning what (not) to forget. In: ECCV. pp. 139–154 (2018)

    2018

  2. [2]

    In: CVPR

    Aljundi, R., Chakravarty, P., Tuytelaars, T.: Expert gate: Lifelong learning with a network of experts. In: CVPR. pp. 3366–3375 (2017)

    2017

  3. [3]

    In: ECCV

    Castro, F.M., Marín-Jiménez, M.J., Guil, N., Schmid, C., Alahari, K.: End-to-end incremental learning. In: ECCV. pp. 233–248 (2018)

    2018

  4. [4]

    arXiv preprint arXiv:1902.10486 (2019)

    Chaudhry, A., Rohrbach, M., Elhoseiny, M., Ajanthan, T., Dokania, P.K., Torr, P.H., Ranzato, M.: On tiny episodic memories in continual learning. arXiv preprint arXiv:1902.10486 (2019)

    Pith/arXiv arXiv 1902

  5. [5]

    In: CVPR

    Chen, D., Krähenbühl, P.: Learning from all vehicles. In: CVPR. pp. 17222–17231 (2022)

    2022

  6. [6]

    In: CoRL

    Chen, D., Zhou, B., Koltun, V., Krähenbühl, P.: Learning by cheating. In: CoRL. pp. 66–75. PMLR (2020)

    2020

  7. [7]

    TPAMI46(12), 10164–10183 (2024)

    Chen, L., Wu, P., Chitta, K., Jaeger, B., Geiger, A., Li, H.: End-to-end autonomous driving: Challenges and frontiers. TPAMI46(12), 10164–10183 (2024)

    2024

  8. [8]

    In: ICCV

    Chitta, K., Prakash, A., Geiger, A.: Neat: Neural attention fields for end-to-end autonomous driving. In: ICCV. pp. 15793–15803 (Oct 2021)

    2021

  9. [9]

    TPAMI45(11), 12878–12895 (2022) 16 J

    Chitta, K., Prakash, A., Jaeger, B., Yu, Z., Renz, K., Geiger, A.: Transfuser: Imita- tion with transformer-based sensor fusion for autonomous driving. TPAMI45(11), 12878–12895 (2022) 16 J. Du, Y. Song et al

    2022

  10. [10]

    Advances in neural information processing systems32(2019)

    De Haan, P., Jayaraman, D., Levine, S.: Causal confusion in imitation learning. Advances in neural information processing systems32(2019)

    2019

  11. [11]

    In: CoRL

    Dosovitskiy, A., Ros, G., Codevilla, F., Lopez, A., Koltun, V.: CARLA: An open urban driving simulator. In: CoRL. pp. 1–16. PMLR (2017)

    2017

  12. [12]

    In: ICCV

    Duan, K., Bai, S., Xie, L., Qi, H., Huang, Q., Tian, Q.: Centernet: Keypoint triplets for object detection. In: ICCV. pp. 6569–6578 (2019)

    2019

  13. [13]

    arXiv preprint arXiv:2503.19755 (2025)

    Fu, H., Zhang, D., Zhao, Z., Cui, J., Liang, D., Zhang, C., Zhang, D., Xie, H., Wang, B., Bai, X.: Orion: A holistic end-to-end autonomous driving framework by vision-language instructed action generation. arXiv preprint arXiv:2503.19755 (2025)

    Pith/arXiv arXiv 2025

  14. [14]

    Communications of the ACM63(11), 139–144 (2020)

    Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM63(11), 139–144 (2020)

    2020

  15. [15]

    In: CVPR

    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: CVPR. pp. 17853– 17862 (2023)

    2023

  16. [16]

    Advances in neural information processing systems26 (2013)

    Hughes, M.C., Sudderth, E.: Memoized online variational inference for dirichlet process mixture models. Advances in neural information processing systems26 (2013)

    2013

  17. [17]

    In: ICCV

    Jaeger, B., Chitta, K., Geiger, A.: Hidden biases of end-to-end driving models. In: ICCV. pp. 8240–8249 (2023)

    2023

  18. [18]

    In: ICCV

    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: ICCV. pp. 7953–7963 (2023)

    2023

  19. [19]

    In: CVPR

    Jia, X., Wu, P., Chen, L., Xie, J., He, C., Yan, J., Li, H.: Think twice before driving: Towards scalable decoders for end-to-end autonomous driving. In: CVPR. pp. 21983–21994 (2023)

    2023

  20. [20]

    In: NeurIPS (2024)

    Jia, X., Yang, Z., Li, Q., Zhang, Z., Yan, J.: Bench2Drive: Towards multi-ability benchmarking of closed-loop end-to-end autonomous driving. In: NeurIPS (2024)

    2024

  21. [21]

    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

  22. [22]

    In: ICCV

    Jiang, B., Chen, S., Xu, Q., Liao, B., Chen, J., Zhou, H., Zhang, Q., Liu, W., Huang,C.,Wang,X.:Vad:Vectorizedscenerepresentationforefficientautonomous driving. In: ICCV. pp. 8340–8350 (2023)

    2023

  23. [23]

    Expert Systems with Applications294, 128739 (2025)

    Jiang, M., Fan, J., Li, F.: Advances in continual learning: A comprehensive review. Expert Systems with Applications294, 128739 (2025)

    2025

  24. [24]

    arXiv preprint arXiv:1711.10563 (2017)

    Kemker, R., Kanan, C.: Fearnet: Brain-inspired model for incremental learning. arXiv preprint arXiv:1711.10563 (2017)

    Pith/arXiv arXiv 2017

  25. [25]

    arXiv preprint arXiv:1312.6114 (2013)

    Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013)

    Pith/arXiv arXiv 2013

  26. [26]

    Proceedings of the national academy of sciences114(13), 3521–3526 (2017)

    Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.: Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences114(13), 3521–3526 (2017)

    2017

  27. [27]

    Journal of Machine Learning Research9(2605), 2579–2605 (2008)

    Laurens, V.D.M., Hinton, G.: Visualizing data using t-sne. Journal of Machine Learning Research9(2605), 2579–2605 (2008)

    2008

  28. [28]

    Journal of Mathematical Psychology91, 128–144 (2019)

    Li, Y., Schofield, E., Gönen, M.: A tutorial on dirichlet process mixture modeling. Journal of Mathematical Psychology91, 128–144 (2019)

    2019

  29. [29]

    TPAMI40(12), 2935–2947 (2017) DeLL 17

    Li, Z., Hoiem, D.: Learning without forgetting. TPAMI40(12), 2935–2947 (2017) DeLL 17

    2017

  30. [30]

    In: CVPR

    Liao, B., Chen, S., Yin, H., Jiang, B., Wang, C., Yan, S., Zhang, X., Li, X., Zhang, Y., Zhang, Q., et al.: Diffusiondrive: Truncated diffusion model for end-to-end autonomous driving. In: CVPR. pp. 12037–12047 (2025)

    2025

  31. [31]

    arXiv preprint arXiv:2508.01158 (2025)

    Lin, Y., Li, Z., Du, G., Zhao, X., Gong, C., Wang, X., Lu, C., Gong, J.: H2c: Hippocampal circuit-inspired continual learning for lifelong trajectory prediction in autonomous driving. arXiv preprint arXiv:2508.01158 (2025)

    Pith/arXiv arXiv 2025

  32. [32]

    Advances in Neural Information Processing Systems36, 44776–44791 (2023)

    Liu, B., Zhu, Y., Gao, C., Feng, Y., Liu, Q., Zhu, Y., Stone, P.: Libero: Benchmark- ing knowledge transfer for lifelong robot learning. Advances in Neural Information Processing Systems36, 44776–44791 (2023)

    2023

  33. [33]

    Advances in neural information processing systems30(2017)

    Lopez-Paz, D., Ranzato, M.: Gradient episodic memory for continual learning. Advances in neural information processing systems30(2017)

    2017

  34. [34]

    Entropy22(11), 1190 (2020)

    Luo, Y., Yin, L., Bai, W., Mao, K.: An appraisal of incremental learning methods. Entropy22(11), 1190 (2020)

    2020

  35. [35]

    In: CVPR

    Mallya, A., Lazebnik, S.: Packnet: Adding multiple tasks to a single network by iterative pruning. In: CVPR. pp. 7765–7773 (2018)

    2018

  36. [36]

    Nature Machine Intelligence pp

    Meng, Y., Bing, Z., Yao, X., Chen, K., Huang, K., Gao, Y., Sun, F., Knoll, A.: Preserving and combining knowledge in robotic lifelong reinforcement learning. Nature Machine Intelligence pp. 1–14 (2025)

    2025

  37. [37]

    Neural Networks113, 54–71 (2019)

    Parisi, G.I., Kemker, R., Part, J.L., Kanan, C., Wermter, S.: Continual lifelong learning with neural networks: A review. Neural Networks113, 54–71 (2019)

    2019

  38. [38]

    John Wiley & Sons (2016)

    Pearl, J., Glymour, M., Jewell, N.P.: Causal inference in statistics: A primer. John Wiley & Sons (2016)

    2016

  39. [39]

    In: CVPR

    Radosavovic, I., Kosaraju, R.P., Girshick, R., He, K., Dollár, P.: Designing network design spaces. In: CVPR. pp. 10428–10436 (June 2020)

    2020

  40. [40]

    In: CVPR (2025)

    Renz, K., Chen, L., Arani, E., Sinavski, O.: Simlingo: Vision-only closed-loop au- tonomous driving with language-action alignment. In: CVPR (2025)

    2025

  41. [41]

    arXiv preprint arXiv:2210.14222 (2022)

    Renz, K., Chitta, K., Mercea, O.B., Koepke, A., Akata, Z., Geiger, A.: Plant: Explainable planning transformers via object-level representations. arXiv preprint arXiv:2210.14222 (2022)

    Pith/arXiv arXiv 2022

  42. [42]

    Columbia University (2024)

    Ruan, K.: When Causality Meets Autonomy: Causal Imitation Learning to Un- ravel Unobserved Influences in Autonomous Driving Decision-Making. Columbia University (2024)

    2024

  43. [43]

    arXiv preprint arXiv:1606.04671 (2016)

    Rusu, A.A., Rabinowitz, N.C., Desjardins, G., Soyer, H., Kirkpatrick, J., Kavukcuoglu, K., Pascanu, R., Hadsell, R.: Progressive neural networks. arXiv preprint arXiv:1606.04671 (2016)

    Pith/arXiv arXiv 2016

  44. [44]

    Shao, H., Wang, L., Chen, R., Li, H., Liu, Y.: Safety-enhanced autonomous driving usinginterpretablesensorfusiontransformer.In:CoRL.pp.726–737.PMLR(2023)

    2023

  45. [45]

    Advances in neural information processing systems30(2017)

    Shin, H., Lee, J.K., Kim, J., Kim, J.: Continual learning with deep generative replay. Advances in neural information processing systems30(2017)

    2017

  46. [46]

    arXiv preprint arXiv:2503.08612 (2025)

    Tang, Y., Xu, Z., Meng, Z., Cheng, E.: Hip-ad: Hierarchical and multi-granularity planning with deformable attention for autonomous driving in a single decoder. arXiv preprint arXiv:2503.08612 (2025)

    Pith/arXiv arXiv 2025

  47. [47]

    Nature Machine Intelligence4(12), 1185–1197 (2022)

    VandeVen,G.M.,Tuytelaars,T.,Tolias,A.S.:Threetypesofincrementallearning. Nature Machine Intelligence4(12), 1185–1197 (2022)

    2022

  48. [48]

    In: CVPR

    Wang, L., He, Z., Dang, R., Shen, M., Liu, C., Chen, Q.: Vision-and-language navigation via causal learning. In: CVPR. pp. 13139–13150 (2024)

    2024

  49. [49]

    arXiv preprint arXiv:2503.12170 (2025) 18 J

    Wang, T., Zhang, C., Qu, X., Li, K., Liu, W., Huang, C.: Diffad: A unified diffu- sion modeling approach for autonomous driving. arXiv preprint arXiv:2503.12170 (2025) 18 J. Du, Y. Song et al

    Pith/arXiv arXiv 2025

  50. [50]

    Advances in Neural Information Processing Systems35, 6119–6132 (2022)

    Wu, P., Jia, X., Chen, L., Yan, J., Li, H., Qiao, Y.: Trajectory-guided control pre- diction for end-to-end autonomous driving: A simple yet strong baseline. Advances in Neural Information Processing Systems35, 6119–6132 (2022)

    2022

  51. [51]

    In: 2024 IEEE Intelligent Vehicles Symposium (IV)

    Yang, Y., Chen, D., Qin, T., Mu, X., Xu, C., Yang, M.: E2e parking: Autonomous parking by the end-to-end neural network on the carla simulator. In: 2024 IEEE Intelligent Vehicles Symposium (IV). pp. 2375–2382. IEEE (2024)

    2024

  52. [52]

    arXiv preprint arXiv:2505.16278 (2025)

    Yang, Z., Chai, Y., Jia, X., Li, Q., Shao, Y., Zhu, X., Su, H., Yan, J.: Drivemoe: Mixture-of-experts for vision-language-action model in end-to-end autonomous driving. arXiv preprint arXiv:2505.16278 (2025)

    Pith/arXiv arXiv 2025

  53. [53]

    arXiv preprint arXiv:2505.17209 (2025)

    Yao, H., Li, P., Jin, B., Zheng, Y., Liu, A., Mu, L., Su, Q., Zhang, Q., Chen, Y., Li, P.: Lilodriver: A lifelong learning framework for closed-loop motion planning in long-tail autonomous driving scenarios. arXiv preprint arXiv:2505.17209 (2025)

    Pith/arXiv arXiv 2025

  54. [54]

    Neural compu- tation31(11), 2266–2291 (2019)

    Yao, X., Huang, T., Wu, C., Zhang, R.X., Sun, L.: Adversarial feature alignment: Avoid catastrophic forgetting in incremental task lifelong learning. Neural compu- tation31(11), 2266–2291 (2019)

    2019

  55. [55]

    arXiv preprint arXiv:2305.10430 (2023)

    Zhai, J.T., Feng, Z., Du, J., Mao, Y., Liu, J.J., Tan, Z., Zhang, Y., Ye, X., Wang, J.: Rethinkingtheopen-loopevaluationofend-to-endautonomousdrivinginnuscenes. arXiv preprint arXiv:2305.10430 (2023)

    Pith/arXiv arXiv 2023

  56. [56]

    In: ICCV

    Zhai, M., Chen, L., Tung, F., He, J., Nawhal, M., Mori, G.: Lifelong gan: Continual learning for conditional image generation. In: ICCV. pp. 2759–2768 (2019)

    2019

  57. [57]

    Baseline + ER

    Zimmerlin, J., Beißwenger, J., Jaeger, B., Geiger, A., Chitta, K.: Hidden biases of end-to-end driving datasets. arXiv preprint arXiv:2412.09602 (2024) DeLL 19 Deconfounded Lifelong Learning for Autonomous Driving via Dynamic Knowledge Spaces Supplementary Material A Comparison with Adaptation of Mainstream Lifelong Learning Methods To the best of our kno...

    Pith/arXiv arXiv 2024

This paper was first reviewed by grok-4.5 on July 14, 2026.