pith. sign in

arxiv: 2606.30940 · v1 · pith:WXDXZFOHnew · submitted 2026-06-29 · 💻 cs.RO · cs.AI· cs.LG

Motion Planning in Compressed Representation Spaces

Lukas Lao Beyer , Sertac Karaman This is my paper

Pith reviewed 2026-07-01 01:06 UTC · model grok-4.3

classification 💻 cs.RO cs.AIcs.LG
keywords motion planningautoencoderlatent space searchdiscrete tokenstrajectory generationroboticsdriving datasetsgenerative models
0
0 comments X

The pith

Motion planning reduces to search in the latent space of hierarchically ordered discrete tokens from a high-compression autoencoder.

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

The paper trains an autoencoder on large motion datasets to represent trajectories as hierarchically ordered discrete tokens with a high compression ratio. Planning then occurs by searching directly inside this token latent space instead of the original high-dimensional space. Because the autoencoder captures the structure of realistic motions, decoded results from the search stay plausible even when the search optimizes an arbitrary objective function supplied only at test time. The approach is shown to support closed-loop planning and multi-agent scenario synthesis on nuPlan and Waymo driving data without task-specific retraining.

Core claim

We learn an autoencoder with a high compression ratio and a latent space of hierarchically ordered, discrete-valued tokens. Leveraging both the dimensionality reduction and the hierarchical coarse-to-fine structure learned by this autoencoder, we then perform motion planning by directly searching in the latent space of tokens. This search can optimize arbitrary objective functions specified at test time, providing a large degree of flexibility while maintaining efficiency and producing realistic solutions by relying on the generative capabilities of the highly compressed autoencoder.

What carries the argument

Autoencoder producing hierarchically ordered discrete tokens in its latent space, used as the representation in which search-based planning occurs.

If this is right

  • Arbitrary objective functions can be optimized at test time without any retraining.
  • Dimensionality reduction and hierarchy keep the search computationally efficient.
  • Realistic trajectories are obtained by decoding from the trained generative autoencoder.
  • The same trained model supports both closed-loop motion planning and multi-agent guided scenario synthesis.

Where Pith is reading between the lines

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

  • The discrete token representation could allow integration with symbolic or combinatorial planning techniques that operate on sequences.
  • Hierarchical ordering suggests a natural way to implement coarse-to-fine refinement where higher-level tokens are fixed first.
  • If similar large-scale trajectory datasets exist in other robotics domains, the same compression-plus-search pattern could be applied without domain-specific redesign.

Load-bearing premise

The latent space of hierarchically ordered discrete tokens learned by the autoencoder preserves enough structure that search inside it produces dynamically feasible, collision-free, and realistic trajectories for arbitrary test-time objective functions.

What would settle it

If trajectories decoded from latent-space search plans frequently violate vehicle dynamics constraints or collide with obstacles when rolled out on the nuPlan and Waymo test sets, the central claim would be falsified.

Figures

Figures reproduced from arXiv: 2606.30940 by Lukas Lao Beyer, Sertac Karaman.

Figure 1
Figure 1. Figure 1: Token search for guided trajectory synthesis. We learn a highly compressed, environment-conditioned autoencoder that maps trajectories to an ordered discrete code z. At test time we perform greedy token search: at step i, we fix z1:i−1, enu￾merate zi, decode each candidate trajectory, score according to the user-specified objective chosen at test time, and keep the best prefix (here we illustrate a goal-re… view at source ↗
Figure 2
Figure 2. Figure 2: Conditional autoencoder training. A conditional autoencoder is trained with a reconstruction objective to capture a highly compact latent representation of the input trajectory given a particular environment. During training, we adaptively inject noise to the latent tokens as a form of soft quantization, and apply causal masking and nested dropout to learn an ordered representation. tion to perform tree se… view at source ↗
Figure 3
Figure 3. Figure 3: Increasing beam size in latent token search with reconstruction objective has diminishing returns. Thanks to the hierarchical structure of the learned latent space, greedy search is already an effective search strategy for reconstruction. (Blue trace: beam search results with Nl = 2; dashed lines: autoencoder reconstructions with and without quantization.) Indeed, [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Tokens have meaningful environment-dependent se￾mantics. When copying the encoding of a given trajectory under a particular environment of the WOMD test set and decoding it under a different test environment (Equation (2)), predictable behavior consistent with the new environment is produced. In (a), Shaded plots show the reference trajectory reconstructed in its original environment, while the remaining p… view at source ↗
Figure 5
Figure 5. Figure 5: Guided trajectory generation examples. Token search synthesizes desired behavior (left turn; green) according to a test-time user-defined objective (left turn, in this example). Compared to the original behavior recorded in the ground truth (straight; blue), the synthesized trajectory is novel in order to comply with the left-turning objective [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Multi-agent token search generates consistent joint trajectories. We generate scenarios (a) and (b) in the same envi￾ronment by performing token search to minimize the deviation be￾tween the final position of the pedestrian (blue) and a user-specified goal point (cross marker). This objective function only supervises the final position of the pedestrian, yet our joint trajectory decoder ensures the behavio… view at source ↗
read the original abstract

Deep learning methods have vastly expanded the capabilities of motion planning in robotics applications, as learning priors from large-scale data has been shown to be essential in capturing the highly complex behavior required for solving tasks such as manipulation or navigation for autonomous vehicles. At the same time, model-based planning algorithms based on search or optimization remain an essential tool due to their flexibility, efficiency, and the ability to incorporate domain knowledge via expert-designed algorithms and objective functions. We propose a new generative framework to unify these two paradigms. First, we learn an autoencoder with a high compression ratio and a latent space of hierarchically ordered, discrete-valued tokens. Leveraging both the dimensionality reduction and the hierarchical coarse-to-fine structure learned by this autoencoder, we then perform motion planning by directly searching in the latent space of tokens. This search can optimize arbitrary objective functions specified at test time, providing a large degree of flexibility while maintaining efficiency and producing realistic solutions by relying on the generative capabilities of the highly compressed autoencoder. We evaluate our method on nuPlan and the Waymo Open Motion Dataset, showing how latent space search can be used for a variety of guided behavior generation tasks, achieving strong performance for closed-loop motion planning and multi-agent guided scenario synthesis without requiring any task-specific training.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The paper proposes a generative framework for motion planning that trains a highly compressed autoencoder whose latent space consists of hierarchically ordered discrete tokens; motion planning is then performed by direct search over these tokens to optimize arbitrary test-time objective functions. The approach is claimed to unify data-driven priors with model-based flexibility, achieving strong performance on closed-loop planning and multi-agent guided scenario synthesis on the nuPlan and Waymo Open Motion Datasets without any task-specific training, by relying on the autoencoder's generative capabilities to produce realistic trajectories.

Significance. If the central claim holds, the method would offer a flexible way to perform test-time optimization of custom objectives while inheriting realism and efficiency from a learned compressed representation, potentially advancing hybrid learning-plus-search approaches in robotics motion planning. The hierarchical token structure and high compression ratio are presented as enabling both coarse-to-fine search and computational efficiency.

major comments (2)
  1. [Abstract] Abstract: the central claim that direct search over the hierarchically ordered discrete tokens 'produces realistic solutions by relying on the generative capabilities of the highly compressed autoencoder' and yields dynamically feasible, collision-free trajectories for arbitrary (unseen) test-time objectives is load-bearing, yet the provided description supplies no mechanism—such as an explicit dynamics model, constraint projection step, feasibility repair, or post-decoding validation—inside the search procedure. Because the autoencoder is trained unsupervised on observed trajectories, nothing in the given account guarantees that out-of-distribution objectives will not decode to infeasible or colliding trajectories; this is the least-secured link in the argument.
  2. [Abstract] Abstract: the evaluation claims 'strong performance for closed-loop motion planning and multi-agent guided scenario synthesis' on nuPlan and Waymo without task-specific training, but supplies no implementation details, baselines, ablation studies, or error analysis. Without these, it is impossible to assess whether the reported results actually support the claim that latent-space search alone suffices for feasibility and realism.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive feedback. We address each major comment below, clarifying the mechanisms described in the full manuscript and indicating where revisions to the abstract will strengthen the presentation.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim that direct search over the hierarchically ordered discrete tokens 'produces realistic solutions by relying on the generative capabilities of the highly compressed autoencoder' and yields dynamically feasible, collision-free trajectories for arbitrary (unseen) test-time objectives is load-bearing, yet the provided description supplies no mechanism—such as an explicit dynamics model, constraint projection step, feasibility repair, or post-decoding validation—inside the search procedure. Because the autoencoder is trained unsupervised on observed trajectories, nothing in the given account guarantees that out-of-distribution objectives will not decode to infeasible or colliding trajectories; this is the least-secured link in the argument.

    Authors: The mechanism is the learned generative prior itself: the autoencoder is trained end-to-end on large-scale real trajectories from nuPlan and Waymo, so its decoder is constrained to map any valid sequence of hierarchical discrete tokens to dynamically feasible, collision-free motions observed in the data distribution. Search occurs entirely inside this compressed latent space; decoded outputs are therefore guaranteed to lie on the manifold of realistic trajectories without an auxiliary dynamics model or repair step. The hierarchical token ordering further supports coarse-to-fine optimization that stays within the learned representation. We agree the abstract would benefit from an explicit sentence on this inductive bias and will revise it accordingly. revision: yes

  2. Referee: [Abstract] Abstract: the evaluation claims 'strong performance for closed-loop motion planning and multi-agent guided scenario synthesis' on nuPlan and Waymo without task-specific training, but supplies no implementation details, baselines, ablation studies, or error analysis. Without these, it is impossible to assess whether the reported results actually support the claim that latent-space search alone suffices for feasibility and realism.

    Authors: The abstract is a high-level summary; the full manuscript (Sections 3–5) supplies the requested details: the autoencoder architecture and training procedure, the exact latent-space search algorithm, closed-loop baselines (including rule-based and learning-based planners), ablation studies on token hierarchy and compression ratio, and quantitative metrics with error analysis on both nuPlan and Waymo. We will revise the abstract to include a concise pointer to these evaluation elements and a summary of the key quantitative findings. revision: partial

Circularity Check

0 steps flagged

No circularity: method is empirical training + search, no derivation reduces to fitted inputs by construction

full rationale

The paper presents an autoencoder trained on trajectory data followed by latent-space search for planning. No equations, uniqueness theorems, or self-citations are shown that would make any claimed output (feasibility, collision avoidance) equivalent to the training inputs by definition. The generative claim is an empirical assertion about the learned model rather than a self-referential derivation. This matches the reader's assessment of score 2.0 with no visible circularity signal.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available; no explicit free parameters, axioms, or invented entities can be extracted beyond the implicit assumption that the autoencoder latent space supports valid planning.

pith-pipeline@v0.9.1-grok · 5748 in / 1098 out tokens · 41777 ms · 2026-07-01T01:06:47.246195+00:00 · methodology

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

47 extracted references · 2 canonical work pages

  1. [1]

    F., Amirloo, E., El-Nouby, A., Zamir, A., and Dehghan, A

    Bachmann, R., Allardice, J., Mizrahi, D., Fini, E., Kar, O. F., Amirloo, E., El-Nouby, A., Zamir, A., and Dehghan, A. FlexTok : Resampling images into 1D token sequences of flexible length. arXiv 2025, 2025

    2025

  2. [2]

    Universal guidance for diffusion models

    Bansal, A., Chu, H.-M., Schwarzschild, A., Sengupta, S., Goldblum, M., Geiping, J., and Goldstein, T. Universal guidance for diffusion models. In Conference on Computer Vision and Pattern Recognition, pp.\ 843--852, 2023

    2023

  3. [3]

    S., Fong, W

    Caesar, H., Kabzan, J., Tan, K. S., Fong, W. K., Wolff, E., Lang, A., Fletcher, L., Beijbom, O., and Omari, S. NuPlan : A closed-loop ML -based planning benchmark for autonomous vehicles. In CVPR ADP3 workshop, 2021

    2021

  4. [4]

    Chang, H., Zhang, H., Jiang, L., Liu, C., and Freeman, W. T. MaskGIT : Masked generative image transformer. In Conference on Computer Vision and Pattern Recognition, June 2022

    2022

  5. [5]

    Rethinking imitation-based planners for autonomous driving

    Cheng, J., Chen, Y., Mei, X., Yang, B., Li, B., and Liu, M. Rethinking imitation-based planners for autonomous driving. In International Conference on Robotics and Automation, pp.\ 14123--14130, 2024

    2024

  6. [6]

    Diffusion policy: Visuomotor policy learning via action diffusion

    Chi, C., Feng, S., Du, Y., Xu, Z., Cousineau, E., Burchfiel, B., and Song, S. Diffusion policy: Visuomotor policy learning via action diffusion. In Robotics: Science and Systems, 2023

    2023

  7. [7]

    VQGAN-CLIP : Open domain image generation and editing with natural language guidance

    Crowson, K., Biderman, S., Kornis, D., Stander, D., Hallahan, E., Castricato, L., and Raff, E. VQGAN-CLIP : Open domain image generation and editing with natural language guidance. arXiv preprint arXiv:2204.08583, 2022

    work page arXiv 2022

  8. [8]

    Vision transformers need registers

    Darcet, T., Oquab, M., Mairal, J., and Bojanowski, P. Vision transformers need registers. In International Conference on Learning Representations, 2024

    2024

  9. [9]

    Parting with misconceptions about learning-based vehicle motion planning

    Dauner, D., Hallgarten, M., Geiger, A., and Chitta, K. Parting with misconceptions about learning-based vehicle motion planning. In Conference on Robot Learning, 2023

    2023

  10. [10]

    and Nichol, A

    Dhariwal, P. and Nichol, A. Diffusion models beat GANs on image synthesis. Advances in Neural Information Processing Systems, 34: 0 8780--8794, 2021

    2021

  11. [11]

    Taming transformers for high-resolution image synthesis

    Esser, P., Rombach, R., and Ommer, B. Taming transformers for high-resolution image synthesis. In Conference on Computer Vision and Pattern Recognition, pp.\ 12873--12883, June 2021

    2021

  12. [12]

    R., Zhou, Y., et al

    Ettinger, S., Cheng, S., Caine, B., Liu, C., Zhao, H., Pradhan, S., Chai, Y., Sapp, B., Qi, C. R., Zhou, Y., et al. Large scale interactive motion forecasting for autonomous driving: The Waymo open motion dataset. In International Conference on Computer Vision, pp.\ 9710--9719, 2021

    2021

  13. [13]

    Time-optimal planning for quadrotor waypoint flight

    Foehn, P., Romero, A., and Scaramuzza, D. Time-optimal planning for quadrotor waypoint flight. Science Robotics, 6 0 (56): 0 eabh1221, 2021

    2021

  14. [14]

    A., and Feron, E

    Frazzoli, E., Dahleh, M. A., and Feron, E. Maneuver-based motion planning for nonlinear systems with symmetries. IEEE Transactions on Robotics, 21 0 (6): 0 1077--1091, 2005

    2005

  15. [15]

    Goh, J. Y. and Gerdes, J. C. Simultaneous stabilization and tracking of basic automobile drifting trajectories. In IEEE Intelligent Vehicles Symposium, pp.\ 597--602, 2016

    2016

  16. [16]

    J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., and Bengio, Y

    Goodfellow, I. J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., and Bengio, Y. Generative adversarial nets. Advances in Neural Information Processing Systems, 27, 2014

    2014

  17. [17]

    and Salimans, T

    Ho, J. and Salimans, T. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021

    2021

  18. [18]

    Denoising diffusion probabilistic models

    Ho, J., Jain, A., and Abbeel, P. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33: 0 6840--6851, 2020

    2020

  19. [19]

    M., Vernaza, P., Phan-Minh, T., Chen, H., Hayden, D

    Huang, X., Wolff, E. M., Vernaza, P., Phan-Minh, T., Chen, H., Hayden, D. S., Edmonds, M., Pierce, B., Chen, X., Jacob, P. E., Chen, X., Tairbekov, C., Agarwal, P., Gao, T., Chai, Y., and Srinivasa, S. Drive GPT : Scaling autoregressive behavior models for driving. In International Conference on Machine Learning, 2025 a

    2025

  20. [20]

    Gameformer: Game-theoretic modeling and learning of transformer-based interactive prediction and planning for autonomous driving

    Huang, Z., Liu, H., and Lv, C. Gameformer: Game-theoretic modeling and learning of transformer-based interactive prediction and planning for autonomous driving. In International Conference on Computer Vision, pp.\ 3903--3913, October 2023

    2023

  21. [21]

    Gen-drive: Enhancing diffusion generative driving policies with reward modeling and reinforcement learning fine-tuning

    Huang, Z., Weng, X., Igl, M., Chen, Y., Cao, Y., Ivanovic, B., Pavone, M., and Lv, C. Gen-drive: Enhancing diffusion generative driving policies with reward modeling and reinforcement learning fine-tuning. In International Conference on Robotics and Automation, pp.\ 3445--3451, 2025 b

    2025

  22. [22]

    MotionDiffuser : Controllable multi-agent motion prediction using diffusion

    Jiang, C., Cornman, A., Park, C., Sapp, B., Zhou, Y., Anguelov, D., et al. MotionDiffuser : Controllable multi-agent motion prediction using diffusion. In Conference on Computer Vision and Pattern Recognition, pp.\ 9644--9653, 2023 a

    2023

  23. [23]

    Efficient planning in a compact latent action space

    Jiang, Z., Zhang, T., Janner, M., Li, Y., Rocktäschel, T., Grefenstette, E., and Tian, Y. Efficient planning in a compact latent action space. International Conference on Learning Representations, 2023 b

    2023

  24. [24]

    Kingma, D. P. and Welling, M. Auto-encoding variational Bayes . In International Conference on Learning Representations , 2014

    2014

  25. [25]

    Highly compressed tokenizer can generate without training

    Lao Beyer, L., Li, T., Chen, X., Karaman, S., and He, K. Highly compressed tokenizer can generate without training. In International Conference on Machine Learning, July 2025

    2025

  26. [26]

    Li, A., Bae, S., Isele, D., Beeson, R., and Tariq, F. M. Predictive planner for autonomous driving with consistency models. arXiv preprint arXiv:2502.08033, 2025

    work page arXiv 2025

  27. [27]

    Autoregressive image generation without vector quantization

    Li, T., Tian, Y., Li, H., Deng, M., and He, K. Autoregressive image generation without vector quantization. Advances in Neural Information Processing Systems, 37: 0 56424--56445, 2024

    2024

  28. [28]

    and Hutter, F

    Loshchilov, I. and Hutter, F. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019

    2019

  29. [29]

    Finite scalar quantization: VQ - VAE made simple

    Mentzer, F., Minnen, D., Agustsson, E., and Tschannen, M. Finite scalar quantization: VQ - VAE made simple. In International Conference on Learning Representations, 2024

    2024

  30. [30]

    One-d-piece: Image tokenizer meets quality-controllable compression

    Miwa, K., Sasaki, K., Arai, H., Takahashi, T., and Yamaguchi, Y. One-d-piece: Image tokenizer meets quality-controllable compression. In ICML Tokenization Workshop, 2025

    2025

  31. [31]

    Robust post-stall perching with a simple fixed-wing glider using lqr-trees

    Moore, J., Cory, R., and Tedrake, R. Robust post-stall perching with a simple fixed-wing glider using lqr-trees. Bioinspiration & biomimetics, 9 0 (2): 0 025013, 2014

    2014

  32. [32]

    Vector quantized models for planning

    Ozair, S., Li, Y., Razavi, A., Antonoglou, I., Van Den Oord, A., and Vinyals, O. Vector quantized models for planning. In International Conference on Machine Learning, pp.\ 8302--8313. PMLR, 2021

    2021

  33. [33]

    StyleCLIP : Text-driven manipulation of stylegan imagery

    Patashnik, O., Wu, Z., Shechtman, E., Cohen-Or, D., and Lischinski, D. StyleCLIP : Text-driven manipulation of stylegan imagery. In International Conference on Computer Vision, pp.\ 2085--2094, October 2021

    2085

  34. [34]

    A., and Kelly, A

    Pivtoraiko, M., Knepper, R. A., and Kelly, A. Differentially constrained mobile robot motion planning in state lattices. Journal of Field Robotics, 26 0 (3): 0 308--333, 2009

    2009

  35. [35]

    R., Su, H., Mo, K., and Guibas, L

    Qi, C. R., Su, H., Mo, K., and Guibas, L. J. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Conference on Computer Vision and Pattern Recognition, pp.\ 652--660, 2017

    2017

  36. [36]

    W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al

    Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pp.\ 8748--8763. PMLR, 2021

    2021

  37. [37]

    Learning ordered representations with nested dropout

    Rippel, O., Gelbart, M., and Adams, R. Learning ordered representations with nested dropout. In International Conference on Machine Learning, pp.\ 1746--1754. PMLR, 2014

    2014

  38. [38]

    High-resolution image synthesis with latent diffusion models

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., and Ommer, B. High-resolution image synthesis with latent diffusion models. In Conference on Computer Vision and Pattern Recognition, pp.\ 10684--10695, 2022

    2022

  39. [39]

    On the pitfalls of heteroscedastic uncertainty estimation with probabilistic neural networks

    Seitzer, M., Tavakoli, A., Antic, D., and Martius, G. On the pitfalls of heteroscedastic uncertainty estimation with probabilistic neural networks. In International Conference on Learning Representations, 2022

    2022

  40. [40]

    Motion transformer with global intention localization and local movement refinement

    Shi, S., Jiang, L., Dai, D., and Schiele, B. Motion transformer with global intention localization and local movement refinement. Advances in Neural Information Processing Systems, 35: 0 6531--6543, 2022

    2022

  41. [41]

    Smith, J. G. The information capacity of amplitude- and variance-constrained scalar Gaussian channels. Information and control, 18 0 (3): 0 203--219, 1971

    1971

  42. [42]

    Loss-guided diffusion models for plug-and-play controllable generation

    Song, J., Zhang, Q., Yin, H., Mardani, M., Liu, M.-Y., Kautz, J., Chen, Y., and Vahdat, A. Loss-guided diffusion models for plug-and-play controllable generation. In International Conference on Machine Learning, volume 202 of Proceedings of Machine Learning Research, pp.\ 32483--32498. PMLR, 23--29 Jul 2023

    2023

  43. [43]

    Neural discrete representation learning

    van den Oord, A., Vinyals, O., and Kavukcuoglu, K. Neural discrete representation learning. In Advances in Neural Information Processing Systems, volume 30, 2017

    2017

  44. [44]

    `` P rincipal components'' enable a new language of images

    Wen, X., Zhao, B., Elezi, I., Deng, J., and Qi, X. `` P rincipal components'' enable a new language of images. In International Conference on Computer Vision, 2025

    2025

  45. [45]

    An image is worth 32 tokens for reconstruction and generation

    Yu, Q., Weber, M., Deng, X., Shen, X., Cremers, D., and Chen, L.-C. An image is worth 32 tokens for reconstruction and generation. In Advances in Neural Information Processing Systems, volume 38, 2024

    2024

  46. [46]

    Carplanner: Consistent auto-regressive trajectory planning for large-scale reinforcement learning in autonomous driving

    Zhang, D., Liang, J., Guo, K., Lu, S., Wang, Q., Xiong, R., Miao, Z., and Wang, Y. Carplanner: Consistent auto-regressive trajectory planning for large-scale reinforcement learning in autonomous driving. In Conference on Computer Vision and Pattern Recognition, pp.\ 17239--17248, 2025

    2025

  47. [47]

    Guided conditional diffusion for controllable traffic simulation

    Zhong, Z., Rempe, D., Xu, D., Chen, Y., Veer, S., Che, T., Ray, B., and Pavone, M. Guided conditional diffusion for controllable traffic simulation. In International Conference on Robotics and Automation, pp.\ 3560--3566, 2023

    2023