Pith. sign in

REVIEW 3 major objections 6 minor 29 references

Mode Collapse Happens: Evaluating Critical Interactions in Joint Trajectory Prediction Models

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A new evaluation framework measures interaction-level mode collapse, mode correctness, and mode coverage in joint trajectory prediction, and finds that learned multimodal models often miss feasible interaction modes in safety-critical…

desk verdict Useful new evaluation framework for interaction-level mode collapse, but the headline numbers rest on a binary homotopy label with a zero threshold and only 41 interaction pairs. read the letter →

arxiv 2506.23164 v1 pith:IGXQVPT6 submitted 2025-06-29 cs.RO cs.AI

Maarten Hugenholtz , Anna Meszaros , Jens Kober , Zlatan Ajanovic This is my paper
classification cs.ROcs.AI
keywords modecollapsejointtrajectorypredictioninteractionmodesfree-endhomotopyinevitablestateautonomousdrivingevaluationnuScenesmultimodal
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Mode collapse in trajectory prediction is usually discussed as a model training failure, but this paper argues that no existing benchmark can actually detect it at the level where it matters: whether a pair of vehicles that will cross paths predicted the yielding interaction correctly. The authors build an evaluation framework that labels each prediction by the homotopy class of the two trajectories, clockwise or counterclockwise rotation, simulates feasible future roll-outs to decide when the interaction mode becomes inevitable, and then reports mode correctness, mode coverage, and mode collapse over only the safety-critical part of each scene. Applying this to four models on nuScenes, the paper finds that AgentFormer leaves 69.8 percent of safety-critical time-steps with at least one feasible interaction mode unpredicted, and that in 9.8 percent of cases the most likely prediction is wrong even at the last moment before the mode becomes inevitable. A constant-velocity extrapolation baseline is often better at predicting the true interaction class than the learned multimodal models.

What carries the argument

The load-bearing tool is free-end homotopy with a two-class label. For each agent pair, the cumulative angular difference $\Delta\theta(\tau^A, \tau^B)$ is thresholded at zero into clockwise (CW) and counterclockwise (CCW) classes, with the static class removed so every trajectory pair receives a definite interaction label. Around this, the framework computes feasible future roll-outs by accelerating one agent while decelerating the other and rejecting colliding pairs with a three-disk collision check, yielding the inevitable homotopy state $t_{h,\mathrm{collapse}} = \min\{t \mid |h_{\mathrm{feas},t}| = 1\}$, the moment only one interaction class remains physically possible. All metrics are evaluated only on the interval from when the ground-truth mode begins to converge until that inevitable state, and mode collapse is defined as a feasible mode missing from the model's K predictions. The time-to-correct-mode-prediction and time-to-covered-mode-prediction metrics convert the same machinery into a temporal measure of when, if ever, the model commits to the right interaction class.

What would settle it

Re-run the evaluation on the same 41 nuScenes interaction pairs with a three-class homotopy label that keeps a static class with a non-zero angle threshold, and compare the resulting mode collapse rates and model rankings; if AgentFormer's 69.8% collapse rate drops sharply or the constant-velocity model no longer leads on mode correctness, the two-class labeling is driving the reported result.

Watch

Extended reading notes

Core claim

The paper's central claim is that interaction-level mode collapse is a real, measurable failure of joint trajectory prediction models, and that it can be quantified with a small set of explicit metrics rather than inferred from distance errors. The authors define a2a mode collapse as a feasible interaction mode that none of the model's K joint predictions covers, and measure it over the interval that starts when the ground-truth homotopy class begins to converge and ends at the inevitable homotopy state. Their results on nuScenes show that the evaluated models do collapse: the strongest multimodal model, AgentFormer, has a mode collapse rate of 69.8 percent, while a unimodal constant-velocity model and a purpose-built oracle model collapse at 100 percent and 18.6 percent respectively. The paper also reports that over a 6-second horizon a simple extrapolation of current velocity predicts the correct interaction mode from the start in 78 percent of cases, outperforming AgentFormer's 56.1 percent, which the authors interpret as evidence that interaction modes are often determined by heading and speed, and that learned diversity does not guarantee interaction diversity.

Load-bearing premise

The framework's numbers all rest on treating every interacting pair as belonging to exactly one of two interaction classes, clockwise or counterclockwise, with the no-rotation class removed; if that binary label misrepresents a scenario, every correctness, coverage, and collapse rate computed from it inherits the error.

Editorial extensions

If this is right

  • Mode collapse is common enough that safety evaluation should track it explicitly: AgentFormer's 69.8% rate means the model frequently fails to output a feasible interaction mode at all.
  • Distance-based leaderboards can rank models opposite to their interaction quality: AgentFormer beats the oracle on joint minADE and minFDE while collapsing far more often, so interaction metrics and displacement metrics should be reported together.
  • A constant-velocity baseline is a strong interaction predictor, so learned multimodal models should be required to beat it on interaction mode correctness, not just on ADE and FDE.
  • With K=5 joint predictions and an interaction mode space that grows exponentially in agent count, full coverage of feasible modes is structurally impossible in dense scenes; the oracle's 18.6% collapse rate is a floor, not a defect.
  • Time-based metrics reveal a safety-relevant limitation: in 9.8% of AgentFormer's 6-second cases the most likely prediction is wrong even at the moment before the interaction becomes inevitable.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper measures presence of modes, not the cost of missing them; combining the same roll-out machinery with a planning cost or a closed-loop simulation would turn a 70% collapse rate into an estimate of collision risk.
  • Replacing the binary CW/CCW threshold with confidence-based or multi-class soft labels would test whether the reported collapse rates are robust to the labeling choice, since any pair of trajectories is currently forced into one of two classes.
  • Because nuScenes contributes only 41 safety-critical interaction pairs in validation, applying the same framework to drone-captured intersection datasets with longer and denser observations would give the time-based metrics enough samples to be stable.
  • The finding that a constant-velocity baseline predicts interaction modes well suggests a cheap training-time sanity metric: monitor whether a learned model's most-likely mode diverges from constant-velocity extrapolation, since those divergences may mark where the model collapses.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes a new evaluation framework for joint trajectory prediction models that focuses on safety-critical interactions between pairs of agents. The framework filters path-crossing interaction pairs, categorizes interaction modes into two classes (CW/CCW) using free-end homotopy with a zero threshold, and computes metrics for mode correctness, mode coverage, and mode collapse, along with time-based variants. The authors apply the framework to four models on nuScenes: AgentFormer (AF), Categorical Traffic Transformer (CTT), a constant-velocity model, and a purpose-built oracle model. The main empirical claim is that mode collapse 'indeed happens': AF shows a mode collapse rate of 69.8% over the safety-critical evaluation interval (Table II), with 9.8% of cases being wrong at the final step before the inevitable homotopy state; the oracle model reduces but does not eliminate collapse (18.6%).

Significance. If the proposed metrics are robust, they address a real gap in trajectory-prediction evaluation: no existing metric explicitly quantifies interaction-level mode collapse, and distance-based metrics such as minADE cannot distinguish between different interaction modes. The temporal metrics (time-to-correct-mode, consistency) are a useful addition. The paper is clearly written, the methodology is computable, and the authors provide code, which supports reproducibility. However, the quantitative claims rest on a small sample (41 interaction pairs, 6 with the full 6-second horizon) and on a binary homotopy threshold whose robustness is not established. The framework is novel and defensible as a proposal, but the specific numerical results should be treated as preliminary until the sensitivity issues are addressed.

major comments (3)
  1. [Section III-B, Eq. (13)] The zero-threshold binary homotopy label assigns every trajectory pair to CW or CCW by the sign of the cumulative angular difference Δθ. For evaluation intervals that begin up to 6 seconds before the inevitable homotopy state, the agents can still be far apart and Δθ can be close to zero; in that regime the class is sensitive to small numerical perturbations and to the choice of θ̂. Since hgt, hml, and hfeas are all derived from this binary label, the headline metrics (e.g., AF mode collapse rate of 69.8% in Table II) are potentially threshold-dependent. Please add a sensitivity analysis varying θ̂ (e.g., θ̂ = 0°, 5°, 10°) and the filtering thresholds dcollision and ΔtPS,max, and report the distribution of |Δθ| over the evaluated time-steps to quantify how often near-zero rotations occur.
  2. [Section III-C, Eqs. (14)-(16)] The feasible-mode set hfeas,t is generated only from two rollouts, (decel, accel) and (accel, decel), as defined in Eq. (14). A 'both decelerate' or 'both stop' outcome, which would yield a near-zero winding angle, is never constructed, so hfeas,t is artificially restricted to at most two classes. The Inevitable Homotopy State (Eq. 17) and the mode collapse rate therefore measure collapse relative to an enumeration that may omit feasible homotopy classes. Please justify why these two acceleration/deceleration profiles are sufficient to capture all feasible homotopy classes for path-crossing interactions, or empirically demonstrate that additional maneuvers (e.g., both decelerate, both stop) leave the set of feasible classes unchanged.
  3. [Section V-B and Table II] The evaluation in Table II is based on only 41 safety-critical interaction pairs in the validation split, of which only 6 allow a full 6-second prediction horizon before the inevitable homotopy state. The paper aggregates the correct/covered/collapse rates over time-steps without reporting confidence intervals or per-pair variance. Given the small sample, numerical differences such as the AF collapse rate of 69.8% versus the oracle's 18.6% could be driven by a few outlier pairs. Please report bootstrap confidence intervals over interaction pairs and provide a per-pair breakdown (e.g., a histogram of collapse rates across pairs) to establish the stability of the reported rates.
minor comments (6)
  1. [Section IV-B] The CTT evaluation is severely compromised: the authors state they could not reproduce the published numbers, that CTT's predictions are effectively unimodal, and that missing agents are filled with static ground-truth positions. This static-fill can create artificial trajectories and biases the interaction metrics. Consider removing CTT from the quantitative comparison or clearly labeling its results as a proof-of-concept rather than a model evaluation.
  2. [Section III-D, Eq. (20)] The definition of th,start is ambiguous if hgt,t never equals hgt,final within the window [th,final - Tp, th,final], for example when the ground-truth homotopy class changes more than once. Please state explicitly how such cases are handled (e.g., fallback to the first available time-step).
  3. [Table II] Please clarify the denominators for the @Tpred and @0s columns: are these percentages over interaction pairs or over time-steps? Also clarify how the mean time-to-correct is computed when a model is correct from the beginning of the interval and no time is available.
  4. [Figure 8] The binned curves in Figure 8 would be more informative if the number of samples per bin were shown, given the small total sample size and the decreasing histogram in Figure 7.
  5. [Conclusion] There is a typo in the concluding paragraph: 'just before it interaction settles' should be 'just before the interaction settles.'
  6. [Introduction] The term 'mode collapse' here is used specifically as failure to cover all feasible interaction modes, which is distinct from the classical GAN-style mode collapse where generated samples collapse to a single mode. Please state this explicitly at first use to avoid confusion with the broader machine-learning literature.

Circularity Check

0 steps flagged · score 0.0 of 10

The paper's evaluation framework is self-contained: its metrics compare model outputs to independently computed ground-truth and feasibility labels, with no load-bearing self-citation or fitted-input-as-prediction step.

full rationale

I walked the derivation chain from Eq. (1) through Eq. (22) and Table II. The homotopy label h (Eq. 13) is computed from observed ground-truth trajectories, not from model outputs. Predicted labels (Eq. 18) are computed from the model's K joint trajectories. The feasible-mode set hfeas,t (Eq. 16) is generated by simulating two collision-checked velocity-profile roll-outs on ground-truth paths and is not a function of the model's outputs. Mode correctness, coverage, and collapse then compare hgt/hfeas against hpred; each comparison uses an independent side of the data. The oracle of Section IV-D is explicitly a privileged baseline constructed to cover feasible modes, and its results are presented as an upper-bound sanity check, not as an empirical prediction; the central 'mode collapse happens' finding rests on AF and CV results, which are not forced by the metric definitions. The threshold choices such as θ̂=0 and the binary CW/CCW categorization are modeling assumptions that affect robustness and interpretation, but they do not make any evaluated quantity equal to its own input. There are also no load-bearing self-citations: prior homotopy and mode-metric work [7], [14] is by other authors and is used as a starting point, not as an unverified uniqueness or existence guarantee. I therefore find no significant circularity.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

No new physical entities are postulated. The free parameters are evaluation thresholds and a modeling choice for the homotopy classification, all of which directly affect the central empirical claim. The most consequential is the zero homotopy threshold, because it defines the binary interaction classes that every metric is computed from.

free parameters (3)
  • dcollision (path-sharing distance threshold) = 1.5 m
    Empirically chosen threshold for defining the path-sharing vector in Eq. (5)-(6). The choice directly determines which agent pairs are considered path-sharing and thus which interactions are classified as safety-critical.
  • dtPS,max (time difference threshold for path-sharing) = 6 s
    Chosen after carefully considering various scenarios from the nuScenes dataset, as stated in Section V-A. This threshold filters the safety-critical interaction set down to 351 pairs from 730.
  • homotopy threshold theta_hat = 0 (static class eliminated)
    The paper sets the threshold to zero to eliminate the static class. This is a modeling choice that affects the definition of all metrics: any nonzero relative angular distance is assigned to CW or CCW, and the resulting binary classification is what all metrics are computed from.
assumptions (3)
  • domain assumption The two-class free-end homotopy label (CW/CCW) is a sufficient and meaningful categorization of interaction modes for safety-critical path-crossing scenarios.
    Invoked in Section III-B when the paper eliminates the static class and reduces all interactions to either CW or CCW. The paper argues this is justified because only safety-critical path-crossing interactions are evaluated, but the binary labeling is still a strong simplification of the variety of interaction behaviors.
  • domain assumption Feasibility of future roll-outs can be determined by a disk-based collision checker and by a fixed set of velocity profile changes (accel/decel) applied to ground truth paths.
    The definition of the inevitable homotopy state and the mode collapse metric depends entirely on this simulation. The paper notes that hypothetical cases can be missed, and it only considers two roll-out profiles per pair, so the feasible set is an approximation of true physical feasibility.
  • domain assumption The ground truth trajectory of each agent is available for the full evaluation window and the interaction pairs can be identified from the data.
    The framework evaluates open-loop predictions against ground truth future trajectories. In an actual deployment the ground truth is unknown, but for offline evaluation this is the standard setup.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Mode Collapse Happens: Evaluating Critical Interactions in Joint Trajectory Prediction Models." pith.science (2026). https://pith.science/paper/IGXQVPT6

@misc{pith2026250623164,
  author       = {Pith},
  title        = {Pith review of: Mode Collapse Happens: Evaluating Critical Interactions in Joint Trajectory Prediction Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IGXQVPT6}},
  note         = {Machine review of arXiv:2506.23164}
}
read the original abstract

Autonomous Vehicle decisions rely on multimodal prediction models that account for multiple route options and the inherent uncertainty in human behavior. However, models can suffer from mode collapse, where only the most likely mode is predicted, posing significant safety risks. While existing methods employ various strategies to generate diverse predictions, they often overlook the diversity in interaction modes among agents. Additionally, traditional metrics for evaluating prediction models are dataset-dependent and do not evaluate inter-agent interactions quantitatively. To our knowledge, none of the existing metrics explicitly evaluates mode collapse. In this paper, we propose a novel evaluation framework that assesses mode collapse in joint trajectory predictions, focusing on safety-critical interactions. We introduce metrics for mode collapse, mode correctness, and coverage, emphasizing the sequential dimension of predictions. By testing four multi-agent trajectory prediction models, we demonstrate that mode collapse indeed happens. When looking at the sequential dimension, although prediction accuracy improves closer to interaction events, there are still cases where the models are unable to predict the correct interaction mode, even just before the interaction mode becomes inevitable. We hope that our framework can help researchers gain new insights and advance the development of more consistent and accurate prediction models, thus enhancing the safety of autonomous driving systems.

Figures

Figures reproduced from arXiv: 2506.23164 by the authors.

Figure 1
Figure 1. We consider an exemplary intersection scenario, with [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Illustration demonstrating the difference between [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 4
Figure 4. Visualization of angular distance calculation and con [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figures from the paper (5 more)
Figure 3
Figure 3. Figure 3: Exemplary traffic scenarios of safety-critical interaction [PITH_FULL_IMAGE:figures/full_fig_p004_3.png]
Figure 5
Figure 5. Figure 5: Visualization of the interaction mode evaluation for AgentFormer on agent-pair (99,2) at frame 11 in scene-0103 of the [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Density heatmap of the path-sharing interactions in [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Histogram of data samples before the inevitable homo [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Relative mode prediction performance plotted against the time to inevitable homotopy state. From left to right, we [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 26 canonical work pages

  1. [1]

    Exploring the implications of autonomous vehicles: a comprehensive review,

    K. Othman, “Exploring the implications of autonomous vehicles: a comprehensive review,” Innovative Infrastructure Solutions , vol. 7, no. 2, p. 165, 2022. [Online]. Available: https://www.ncbi.nlm.nih.gov/ pmc/articles/PMC8885781/

    work page 2022

  2. [2]

    The integration of prediction and planning in deep learning automated driving systems: A review,

    S. Hagedorn, M. Hallgarten, M. Stoll, and A. P. Condurache, “The integration of prediction and planning in deep learning automated driving systems: A review,”IEEE Transactions on Intelligent Vehicles, pp. 1–17, 2024

    work page 2024

  3. [3]

    Social GAN: Socially Acceptable Trajectories with Generative Adversarial Networks,

    A. Gupta, J. Johnson, L. Fei-Fei, S. Savarese, and A. Alahi, “Social GAN: Socially Acceptable Trajectories with Generative Adversarial Networks,” in 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Salt Lake City, UT: IEEE, Jun. 2018, pp. 2255– 2264

    work page 2018

  4. [4]

    Social Ways: Learning Multi- Modal Distributions of Pedestrian Trajectories With GANs,

    J. Amirian, J.-B. Hayet, and J. Pettre, “Social Ways: Learning Multi- Modal Distributions of Pedestrian Trajectories With GANs,” in 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). Long Beach, CA, USA: IEEE, Jun. 2019, pp. 2964–2972

    work page 2019

  5. [5]

    TNT: Target- driven Trajectory Prediction,

    H. Zhao, J. Gao, T. Lan, C. Sun, B. Sapp, B. Varadarajan, Y . Shen, Y . Shen, Y . Chai, C. Schmid, C. Li, and D. Anguelov, “TNT: Target- driven Trajectory Prediction,” in Proceedings of the 2020 Conference on Robot Learning. PMLR, Oct. 2021, pp. 895–904, iSSN: 2640-3498

    work page 2020

  6. [6]

    DenseTNT: End-to-end Trajectory Prediction from Dense Goal Sets,

    J. Gu, C. Sun, and H. Zhao, “DenseTNT: End-to-end Trajectory Prediction from Dense Goal Sets,” in 2021 IEEE/CVF International Conference on Computer Vision (ICCV). Montreal, QC, Canada: IEEE, Oct. 2021, pp. 15 283–15 292

    work page 2021

  7. [7]

    Categorical Traffic Transformer: Interpretable and Diverse Behavior Prediction with Tokenized Latent,

    Y . Chen, S. Tonkens, and M. Pavone, “Categorical Traffic Transformer: Interpretable and Diverse Behavior Prediction with Tokenized Latent,” Nov. 2023, arXiv:2311.18307 [cs]

    arXiv 2023

  8. [8]

    ScePT: Scene-consistent, Policy- based Trajectory Predictions for Planning,

    Y . Chen, B. Ivanovic, and M. Pavone, “ScePT: Scene-consistent, Policy- based Trajectory Predictions for Planning,” in 2022 IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR) . New Orleans, LA, USA: IEEE, Jun. 2022, pp. 17 082–17 091. JOURNAL OF LATEX CLASS FILES, VOL. X, NO. X, APRIL 2025 12

    work page 2022

Show all 29 references
  1. [9]

    DLow: Diversifying Latent Flows for Di- verse Human Motion Prediction,

    Y . Yuan and K. Kitani, “DLow: Diversifying Latent Flows for Di- verse Human Motion Prediction,” in Computer Vision – ECCV 2020 , A. Vedaldi, H. Bischof, T. Brox, and J.-M. Frahm, Eds. Cham: Springer International Publishing, 2020, pp. 346–364

    2020

  2. [10]

    Trajectory Forecasts in Unknown Environ- ments Conditioned on Grid-Based Plans,

    N. Deo and M. M. Trivedi, “Trajectory Forecasts in Unknown Environ- ments Conditioned on Grid-Based Plans,” Apr. 2021, arXiv:2001.00735 [cs]. [Online]. Available: http://arxiv.org/abs/2001.00735

    2021 arXiv

  3. [11]

    Search- Based Optimal Motion Planning for Automated Driving,

    Z. Ajanovi ´c, B. Lacevic, B. Shyrokau, M. Stolz, and M. Horn, “Search- Based Optimal Motion Planning for Automated Driving,” Oct. 2018

    2018

  4. [12]

    Feasible and Adaptive Multimodal Trajectory Prediction with Semantic Maneuver Fusion,

    H. Berkemeyer, R. Franceschini, T. Tran, L. Che, and G. Pipa, “Feasible and Adaptive Multimodal Trajectory Prediction with Semantic Maneuver Fusion,” in 2021 IEEE International Conference on Robotics and Automation (ICRA), May 2021, pp. 8530–8536, iSSN: 2577-087X

    2021

  5. [13]

    Interaction-Based Trajectory Prediction Over a Hybrid Traffic Graph,

    S. Kumar, Y . Gu, J. Hoang, G. C. Haynes, and M. Marchetti-Bowick, “Interaction-Based Trajectory Prediction Over a Hybrid Traffic Graph,” in 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Sep. 2021, pp. 5530–5535, iSSN: 2153-0866

    2021

  6. [14]

    Interactive Joint Planning for Autonomous Vehicles,

    Y . Chen, S. Veer, P. Karkus, and M. Pavone, “Interactive Joint Planning for Autonomous Vehicles,” IEEE Robotics and Automation Letters , vol. 9, no. 2, pp. 987–994, Feb. 2024, conference Name: IEEE Robotics and Automation Letters

    2024

  7. [15]

    Tree- structured Policy Planning with Learned Behavior Models,

    Y . Chen, P. Karkus, B. Ivanovic, X. Weng, and M. Pavone, “Tree- structured Policy Planning with Learned Behavior Models,” in 2023 IEEE International Conference on Robotics and Automation (ICRA) , May 2023, pp. 7902–7908

    2023

  8. [16]

    JFP: Joint Future Prediction with Interactive Multi-Agent Modeling for Autonomous Driving,

    W. Luo, C. Park, A. Cornman, B. Sapp, and D. Anguelov, “JFP: Joint Future Prediction with Interactive Multi-Agent Modeling for Autonomous Driving,” in Proceedings of The 6th Conference on Robot Learning. PMLR, Mar. 2023, pp. 1457–1467, iSSN: 2640-3498. [Online]. Available: htt...

    2023

  9. [17]

    AgentFormer: Agent- Aware Transformers for Socio-Temporal Multi-Agent Forecasting,

    Y . Yuan, X. Weng, Y . Ou, and K. Kitani, “AgentFormer: Agent- Aware Transformers for Socio-Temporal Multi-Agent Forecasting,” in 2021 IEEE/CVF International Conference on Computer Vision (ICCV) . Montreal, QC, Canada: IEEE, Oct. 2021, pp. 9793–9803

    2021

  10. [18]

    Large Scale Interactive Motion Forecasting for Autonomous Driving : The Waymo Open Motion Dataset,

    S. Ettinger, S. Cheng, B. Caine, C. Liu, H. Zhao, S. Pradhan, Y . Chai, B. Sapp, C. Qi, Y . Zhou, Z. Yang, A. Chouard, P. Sun, J. Ngiam, V . Vasudevan, A. McCauley, J. Shlens, and D. Anguelov, “Large Scale Interactive Motion Forecasting for Autonomous Driving : The Waymo Open ...

    2021

  11. [19]

    Defining interactions: a conceptual framework for understanding interactive behaviour in human and automated road traffic,

    G. Markkula, R. Madigan, D. Nathanael, E. Portouli, Y . M. Lee, A. Diet- rich, J. Billington, A. Schieben, and N. Merat, “Defining interactions: a conceptual framework for understanding interactive behaviour in human and automated road traffic,” Theoretical Issues in Ergonomic...

    2020

  12. [20]

    Topological constraints in search-based robot path planning,

    S. Bhattacharya, M. Likhachev, and V . Kumar, “Topological constraints in search-based robot path planning,” Autonomous Robots, vol. 3, no. 33, pp. 273–290, 2012

    2012

  13. [21]

    Fast collision checking for intelligent vehicle motion planning,

    J. Ziegler and C. Stiller, “Fast collision checking for intelligent vehicle motion planning,” in 2010 IEEE Intelligent Vehicles Symposium . La Jolla, CA, USA: IEEE, Jun. 2010, pp. 518–522

    2010

  14. [22]

    nuScenes: A Multi- modal Dataset for Autonomous Driving,

    H. Caesar, V . Bankiti, A. H. Lang, S. V ora, V . E. Liong, Q. Xu, A. Krishnan, Y . Pan, G. Baldan, and O. Beijbom, “nuScenes: A Multi- modal Dataset for Autonomous Driving,” in 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , Jun. 2020, pp. 11 618–...

    2020

  15. [23]

    Khrylx/AgentFormer,

    Y . Yuan, “Khrylx/AgentFormer,” May 2024, original-date: 2021- 03-24T16:40:46Z. [Online]. Available: https://github.com/Khrylx/ AgentFormer

    2024

  16. [24]

    NVlabs/diffstack at CTT release

    “NVlabs/diffstack at CTT release.” [Online]. Available: https://github. com/NVlabs/diffstack/tree/CTT release

  17. [25]

    Scenario Under- standing and Motion Prediction for Autonomous Vehicles - Review and Comparison,

    P. Karle, M. Geisslinger, J. Betz, and M. Lienkamp, “Scenario Under- standing and Motion Prediction for Autonomous Vehicles - Review and Comparison,” IEEE Transactions on Intelligent Transportation Systems , vol. 23, no. 10, pp. 16 962–16 982, 2022

    2022

  18. [26]

    Standards for passenger comfort in automated vehicles: Acceleration and jerk,

    K. N. de Winkel, T. Irmak, R. Happee, and B. Shyrokau, “Standards for passenger comfort in automated vehicles: Acceleration and jerk,” Applied Ergonomics, vol. 106, p. 103881, Jan. 2023

    2023

  19. [27]

    The rounD Dataset: A Drone Dataset of Road User Trajectories at Roundabouts in Germany,

    R. Krajewski, T. Moers, J. Bock, L. Vater, and L. Eckstein, “The rounD Dataset: A Drone Dataset of Road User Trajectories at Roundabouts in Germany,” in 2020 IEEE 23rd International Conference on Intelligent Transportation Systems (ITSC), Sep. 2020, pp. 1–6. [Online]. Availabl...

    2020

  20. [28]

    The inD Dataset: A Drone Dataset of Naturalistic Road User Tra- jectories at German Intersections,

    J. Bock, R. Krajewski, T. Moers, S. Runde, L. Vater, and L. Eckstein, “The inD Dataset: A Drone Dataset of Naturalistic Road User Tra- jectories at German Intersections,” in 2020 IEEE Intelligent Vehicles Symposium (IV), Oct. 2020, pp. 1929–1934, iSSN: 2642-7214

    2020

  21. [29]

    INTERACTION Dataset: An INTERnational, Adversarial and Cooper- ative moTION Dataset in Interactive Driving Scenarios with Semantic Maps,

    W. Zhan, L. Sun, D. Wang, H. Shi, A. Clausse, M. Naumann, J. Kum- merle, H. Konigshof, C. Stiller, A. de La Fortelle, and M. Tomizuka, “INTERACTION Dataset: An INTERnational, Adversarial and Cooper- ative moTION Dataset in Interactive Driving Scenarios with Semantic Maps,” Sep...

    2019 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.