Pith. sign in

REVIEW 4 major objections 5 minor 48 references

Foresight in Motion: Reinforcing Trajectory Prediction with Reward Heuristics

T0 review · 4 major / 5 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 The paper argues that trajectory prediction improves when a model first recovers a driver's reward field from past trajectories by inverse reinforcement learning, samples plausible intention paths from that reward, and only then decodes a…

desk verdict A genuinely new IRL-based reasoning module with competitive Argoverse 1 results, but the headline SOTA claim rests on a customized Argoverse 2 benchmark that isn't a fair comparison. read the letter →

arxiv 2507.12083 v1 pith:G2LCXJZY submitted 2025-07-16 cs.CV cs.RO

Muleilan Pei , Shaoshuai Shi , Xuesong Chen , Xu Liu , Shaojie Shen This is my paper
classification cs.CVcs.RO
keywords trajectorypredictionmotionforecastinginversereinforcementlearningintentionreasoningquery-centricencodingstatespacemodelsMambaautonomousdriving
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

This paper tries to establish that trajectory forecasting for autonomous driving improves when the model thinks like a planner instead of a regressor: first infer what the driver might be trying to do, then predict where they will go. The proposed system, FiM, learns a reward field over a bird's-eye-view grid from past trajectories using maximum-entropy inverse reinforcement learning, rolls that reward out into many candidate intention sequences, and feeds those intentions as spatial priors into a decoder that emits multimodal trajectories with probabilities. The authors claim this strategy outperforms direct end-to-end predictors on the Argoverse and nuScenes benchmarks, with the clearest gains in confidence-sensitive metrics such as the Brier score. This matters because an autonomous vehicle's downstream planner needs calibrated probabilities, not just a low average error, to decide how to react to each other road user.

What carries the argument

The load-bearing machinery is the reward-driven intention reasoner built on QIRL, a query-centric adaptation of maximum-entropy inverse reinforcement learning. The bird's-eye-view grid serves as the state space of a Markov decision process; each cell's learned query token is pushed through $1\times1$ convolutions to produce a scalar reward, and future trajectories quantized to grid resolution $d$ act as expert demonstrations. MaxEnt IRL iterates between updating the reward and solving the inner forward reinforcement-learning problem until the loss converges, after which parallel rollouts over the induced policy yield $L$ GRTs, sequences of intended cells with their associated tokens, that carry intention information into the decoder. The second piece is the hierarchical DETR-like decoder with Bi-Mamba: an anchor-free proposal query attends to the reasoning tokens to generate proposals, K-means clustering reduces them to $K$ anchors, a refinement stage re-encodes the anchors, and a bidirectional selective state-space model with two classification tokens aggregates the sequential trajectory features for final regression and mode classification. An auxiliary spatial-temporal occupancy grid head densifies the grid features by predicting future agent occupancy.

What would settle it

A decisive experiment is a reward-substitution ablation: keep the entire pipeline fixed but replace the IRL-derived reward with a constant field or a hand-coded distance-to-centerline cost before rolling out the intention paths; if the reported minFDE and Brier-score gains over the vanilla baseline persist under the substitute reward, then the learned reward itself is not what carries the improvement, and the reasoning is just extra conditioning.

Watch

Extended reading notes

Core claim

The central claim is that explicit, interpretable intention reasoning, recovered as a reward distribution rather than hand-crafted rules, is a learnable prior that makes trajectory prediction simultaneously more accurate and more confident. In the proposed pipeline, scene context is aggregated into grid-shaped queries; a query-centric IRL module (QIRL) maps those queries to a per-cell reward field, with quantized future trajectories serving as expert demonstrations for maximum-entropy IRL. Policy rollouts over the reward sample $L$ grid-based reasoning traversals (GRTs), long-horizon intention sequences, and the sampled cell tokens become priors for a hierarchical DETR-like decoder whose bidirectional Mamba refinement reads the trajectory sequence forward and backward to produce $K$ final trajectories with probabilities. The paper reports that this design posts the best single-model Brier score among the listed Argoverse 1 entries, the lowest minADE5 (0.88) and MR10 (0.23) among the listed nuScenes entries, and outperforms QCNet and DeMo on a customized Argoverse 2 benchmark, where extending the intention-supervision horizon from 30 to 60 timestamps monotonically improves confidence metrics. The message is that reasoning about where an agent could plausibly go, guided by a learned reward, does real work that direct regression leaves on the table.

Load-bearing premise

The load-bearing premise is that trajectories snapped onto a coarse prediction grid can serve as expert demonstrations for inverse reinforcement learning and yield a reward field that is compact enough to reason over yet faithful enough to capture a driver's true intentions, and that this reward transfers from training scenes to unseen ones so the sampled intention priors guide rather than corrupt the final forecasts.

Editorial extensions

If this is right

  • If the central claim holds, the reward map and the sampled intention paths give the prediction an interpretable intermediate: a human could inspect why the model favors certain futures, addressing the black-box critique of direct forecasting.
  • Longer intention-supervision horizons monotonically improve confidence metrics in the Argoverse 2 experiments, implying that prediction confidence can be purchased by deeper reasoning rather than by larger decoders.
  • Because QIRL accepts either trajectories or long-range paths as demonstration input, the same architecture can exploit long-horizon data when available at deployment without architectural changes.
  • The consistent gains across Argoverse 1, Argoverse 2, and nuScenes indicate the reasoning prior transfers across scene types, city layouts, and sensor configurations.
  • The auxiliary occupancy-grid head also improves accuracy when added, suggesting that modeling future agent interactions compounds with intention reasoning.

Reading between the lines

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

  • A natural extension the paper does not run: replace the IRL-learned reward with a hand-designed potential field or a constant field while freezing the decoder; if the gains survive, the specific learned reward is not the operative ingredient, and the mechanism is just the extra conditioning signal.
  • The confidence gains in the Brier score suggest a safety application the authors leave implicit: the predicted mode probabilities could directly gate a downstream planner's aggressiveness, suppressing low-confidence maneuvers.
  • The grid resolution $d$ is a free parameter whose effect on the IRL transfer is never swept; a testable prediction is that gains degrade sharply once $d$ exceeds roughly a lane width, since the reward would then lose the spatial precision the final trajectory needs.
  • The long-horizon confidence effect was demonstrated on a customized Argoverse 2 protocol rather than a standard leaderboard; confirming GRT-L on the standard benchmark would settle whether the effect survives official evaluation conditions.
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

4 major / 5 minor

Summary. The paper proposes FiM, a trajectory forecasting model that first reasons about an agent's future intentions by learning a reward distribution through a query-centric maximum-entropy inverse reinforcement learning (QIRL) scheme over a discretized BEV grid, then uses policy rollouts (called Grid-based Reasoning Traversals, GRTs) as priors for a hierarchical DETR-like decoder augmented with a bidirectional Mamba architecture and an auxiliary spatial-temporal occupancy grid map head. The method is evaluated on Argoverse 1, a customized Argoverse 2 validation protocol, and nuScenes, reporting competitive accuracy and improved confidence metrics (Brier score, brier-minFDE). The central claim is that this 'First Reasoning, Then Forecasting' strategy improves both prediction accuracy and confidence and outperforms state-of-the-art models.

Significance. If the results are reproducible and the comparisons are fair, the paper makes a useful contribution: it demonstrates a concrete way to inject planning-style intention reasoning into a modern vectorized motion forecasting pipeline, and the ablation suite (Tables 4-8) provides evidence that the reasoning module, the OGM head, and the Bi-Mamba decoder each contribute to the final performance. The emphasis on confidence calibration through Brier score is practically relevant for downstream planning and decision-making. The main caveats are that the headline Argoverse 2 result is obtained on a customized benchmark with auxiliary future supervision unavailable to the compared baselines, and the nuScenes claim of outperforming all leaderboard entries is not backed by a full leaderboard comparison. These issues directly affect the strength of the central claim but are fixable with additional experiments and more careful reporting.

major comments (4)
  1. [Sec. 4.2, Table 2] The Argoverse 2 comparison is not apples-to-apples. FiM's GRT-S, GRT-M, and GRT-L variants are trained with auxiliary intention supervision from future timestamps beyond the 30-step prediction horizon (30, 45, or 60 future positions, respectively), while QCNet and DeMo are evaluated using their official implementations, which are trained on the standard 60-step forecasting task without such auxiliary future-path supervision. Under this customized protocol, the reported improvements (e.g., Brier score 0.603 vs. 0.626) may be attributable to the extra future supervision rather than to the proposed reasoning module. The paper should either report FiM on the standard Argoverse 2 validation split (60-step, no auxiliary future supervision) or retrain all baselines on the same customized task; without this, the claim of outperforming state-of-the-art models on Argoverse 2 is unsupported.
  2. [Sec. 4.2, Table 3 and surrounding text] The statement that FiM 'outperforms all current entries on the leaderboard' on nuScenes is not substantiated. Table 3 lists only seven prior methods and does not provide a complete leaderboard snapshot, submission identifiers, or a date-stamped list of all entries. Several published nuScenes forecasting methods are not included, and no official leaderboard run is cited. The claim should either be restricted to 'outperforms the baselines listed in Table 3' or supported by a complete, dated leaderboard comparison.
  3. [Sec. 3.4, QIRL] The core assumption that the learned reward distribution is a 'compact yet informative representation' of the target agent's behavior and that policy rollouts produce useful priors is not tested with respect to the free parameters d, H, W, and L. No ablation varies the grid resolution d or the number of rollouts L, and the only direct evidence for the reasoning module is Table 4, which compares against a cross-attention alternative. Because the method's contribution depends on the reward heuristic surviving discretization and transferring to unseen scenes, the paper should include sensitivity analysis over d and L, and ideally an analysis of reward generalizability.
  4. [Appendix references and reproducibility] The manuscript repeatedly references 'Appx. 6.1–6.3' and 'Appx. 7' for the MaxEnt IRL formulation, decoder details, training hyperparameters, and additional qualitative results, but the submitted version as provided contains no appendix. Without these details, the central QIRL training objective L_IRL and the rollout procedure are not fully specified, preventing verification and reproduction of the reported results. The missing appendix should be included, or the key equations and hyperparameters should be moved into the main text.
minor comments (5)
  1. [Abstract and Sec. 1] The abstract and conclusion describe the performance as 'highly competitive' while contribution (4) states 'outperforming other state-of-the-art models'; these statements should be aligned once the benchmark comparison issues are resolved.
  2. [Sec. 4.1, Table 1] The experimental settings do not describe the ensembling procedure used for the lower group of Table 1, so the relationship between the single-model and ensemble rows is unclear.
  3. [Sec. 4.2, Table 2] For GRT-M and GRT-L, the paper states that reasoning modules are trained with 45 and 60 future timestamps, but it is not explicitly stated whether the trajectory decoder for these variants is still supervised only on the first 30 timestamps; please clarify the exact supervision split for each row.
  4. [Sec. 3.5] The K-means clustering of trajectory proposals is mentioned without specifying the distance metric, the number of iterations, or how the number of clusters K is determined; please add these implementation details.
  5. [Eq. (2)] The overall loss includes L_IRL, but the precise mathematical form of L_IRL is only deferred to an appendix that is not present in the submitted version; this should be fixed by including the appendix or stating the loss explicitly in the main text.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the derivation is a standard supervised pipeline and the central claims are evaluated against external benchmarks.

full rationale

The paper's derivation chain contains no step in which a claimed prediction is equivalent to its input by construction. The reward distribution in Sec. 3.4 is learned from quantized ground-truth future trajectories via MaxEnt IRL, and the final trajectory decoder in Sec. 3.6 is trained with Huber loss against the same ground-truth future positions; this is ordinary supervised training, not a fitted parameter renamed as a prediction. The GRT rollouts used at inference are generated from a learned reward and do not consume test-time ground truth. Self-citations such as GoIRL [27] supply technical details for the IRL implementation, but the core IRL formulation is grounded in the external MaxEnt IRL work [48], and the main performance claims are supported by external Argoverse 1 and nuScenes leaderboard comparisons. The customized Argoverse 2 protocol gives FiM additional auxiliary future-path supervision during training, which raises a comparison-fairness concern, but that is an evaluation-design issue rather than circular reasoning. Overall, the central derivation is self-contained with respect to its inputs and no circular step was identified.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The paper introduces new model components (QIRL, GRT, dual-mode Bi-Mamba, S-T OGM head) but no new physical entities, forces, or conserved quantities. These components are empirically validated on standard benchmarks, so the 'graviton problem' does not apply.

free parameters (6)
  • Grid resolution d = not reported in main text
    Defines real-world extent of each grid cell for reward computation and GRT quantization; chosen by hand and central to the reasoning module.
  • Grid dimensions H, W = not reported
    Spatial coverage of the BEV grid used for reward and OGM heads; hand-set.
  • Number of rollouts L = not reported
    Number of GRT samples used as intention priors; set to L >> K.
  • GRT planning horizon = 30/45/60 in Argoverse 2 variants; not stated for other datasets
    Length of the intention sequence; controls how much future information is captured.
  • Loss weights alpha, beta, gamma = not reported
    Balance IRL, OGM, regression, and classification losses; hand-tuned.
  • Number of modes K = 6 in experiments
    Number of output trajectories and anchor modes; standard for benchmarks.
assumptions (6)
  • standard math Maximum Entropy IRL (Ziebart et al. 2008) can recover a valid reward function from expert demonstrations and is differentiable enough to be integrated end-to-end.
    Sec. 3.4 builds the QIRL module directly on this method; the paper relies on its convergence and stability without proof.
  • domain assumption Road users behave rationally, so a reward-based planning view can model their future intentions.
    State in the Introduction: 'under the assumption that road users behave rationally'.
  • domain assumption Quantizing future trajectories to a grid at resolution d preserves the information required for intention reasoning and downstream decoding.
    Sec. 3.4: 'Future trajectories are quantized to a resolution d to form expert demonstration states'. If this discretization loses critical precision, the reward and GRT priors are degraded.
  • ad hoc to paper The learned reward distribution is a compact and sufficient representation of the target agent's intention for guiding trajectory generation.
    Sec. 3.4: 'The learned reward serves as a compact representation, capturing observed behaviors and underlying intentions of the agent.' This is the core modeling bet and is not derived.
  • domain assumption Availability of future trajectories beyond the prediction horizon during training (for intention supervision) is a reasonable assumption for real applications.
    Sec. 4.2 Argoverse 2 setup: models are trained with the subsequent 30 positions as auxiliary intention supervision; this is an assumption about data availability in practice.
  • standard math Selective state-space models (Mamba) are suitable for modeling sequential dependencies in trajectory tokens.
    Sec. 3.5 adopts Bi-Mamba based on prior success; no new justification is given beyond citations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Foresight in Motion: Reinforcing Trajectory Prediction with Reward Heuristics." pith.science (2026). https://pith.science/paper/G2LCXJZY

@misc{pith2026250712083,
  author       = {Pith},
  title        = {Pith review of: Foresight in Motion: Reinforcing Trajectory Prediction with Reward Heuristics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/G2LCXJZY}},
  note         = {Machine review of arXiv:2507.12083}
}
read the original abstract

Motion forecasting for on-road traffic agents presents both a significant challenge and a critical necessity for ensuring safety in autonomous driving systems. In contrast to most existing data-driven approaches that directly predict future trajectories, we rethink this task from a planning perspective, advocating a "First Reasoning, Then Forecasting" strategy that explicitly incorporates behavior intentions as spatial guidance for trajectory prediction. To achieve this, we introduce an interpretable, reward-driven intention reasoner grounded in a novel query-centric Inverse Reinforcement Learning (IRL) scheme. Our method first encodes traffic agents and scene elements into a unified vectorized representation, then aggregates contextual features through a query-centric paradigm. This enables the derivation of a reward distribution, a compact yet informative representation of the target agent's behavior within the given scene context via IRL. Guided by this reward heuristic, we perform policy rollouts to reason about multiple plausible intentions, providing valuable priors for subsequent trajectory generation. Finally, we develop a hierarchical DETR-like decoder integrated with bidirectional selective state space models to produce accurate future trajectories along with their associated probabilities. Extensive experiments on the large-scale Argoverse and nuScenes motion forecasting datasets demonstrate that our approach significantly enhances trajectory prediction confidence, achieving highly competitive performance relative to state-of-the-art methods.

Figures

Figures reproduced from arXiv: 2507.12083 by the authors.

Figure 1
Figure 1. Comparison between the conventional motion forecast [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed framework. Our approach integrates a reward-driven intention reasoner into an encoder-decoder [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The process of Bi-Mamba-enhanced decoding. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Qualitative results of our model on the Argoverse validation set. The historical trajectory, ground-truth future trajectory, and mul [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

48 extracted references · 36 canonical work pages

  1. [1]

    nuscenes: A multimodal dataset for autonomous driving

    Holger Caesar, Varun Bankiti, Alex H Lang, et al. nuscenes: A multimodal dataset for autonomous driving. In Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621–11631, 2020. 6

    work page 2020

  2. [2]

    End-to- end object detection with transformers

    Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to- end object detection with transformers. In European confer- ence on computer vision, pages 213–229. Springer, 2020. 4

    work page 2020

  3. [3]

    Multipath: Multiple probabilistic anchor tra- jectory hypotheses for behavior prediction

    Yuning Chai, Benjamin Sapp, Mayank Bansal, and Dragomir Anguelov. Multipath: Multiple probabilistic anchor tra- jectory hypotheses for behavior prediction. arXiv preprint arXiv:1910.05449, 2019. 2

    arXiv 1910

  4. [4]

    Argoverse: 3d tracking and forecasting with rich maps

    Ming-Fang Chang, John Lambert, et al. Argoverse: 3d tracking and forecasting with rich maps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8748–8757, 2019. 6

    work page 2019

  5. [5]

    Multimodal trajectory predictions for autonomous driving using deep convolutional networks

    Henggang Cui, Vladan Radosavljevic, Fang-Chieh Chou, Tsung-Han Lin, Thi Nguyen, Tzu-Kuo Huang, Jeff Schnei- der, and Nemanja Djuric. Multimodal trajectory predictions for autonomous driving using deep convolutional networks. In 2019 international conference on robotics and automation (icra), pages 2090–2096. IEEE, 2019. 2

    work page 2019

  6. [6]

    Trajectory forecasts in unknown environments conditioned on grid-based plans

    Nachiket Deo and Mohan M Trivedi. Trajectory forecasts in unknown environments conditioned on grid-based plans. arXiv preprint arXiv:2001.00735, 2020. 2, 6

    arXiv 2001

  7. [7]

    Multimodal trajectory prediction conditioned on lane-graph traversals

    Nachiket Deo, Eric Wolff, and Oscar Beijbom. Multimodal trajectory prediction conditioned on lane-graph traversals. In Conference on Robot Learning, pages 203–212, 2021. 1, 6

    work page 2021

  8. [8]

    Mac- former: Map-agent coupled transformer for real-time and ro- bust trajectory prediction

    Chen Feng, Hangning Zhou, Huadong Lin, Zhigang Zhang, Ziyao Xu, Chi Zhang, Boyu Zhou, and Shaojie Shen. Mac- former: Map-agent coupled transformer for real-time and ro- bust trajectory prediction. IEEE Robotics and Automation Letters, 2023. 6

    work page 2023

Show all 48 references
  1. [9]

    Unitraj: A unified framework for scalable vehi- cle trajectory prediction

    Lan Feng, Mohammadhossein Bahari, Kaouther Mes- saoud Ben Amor, ´Eloi Zablocki, Matthieu Cord, and Alexan- dre Alahi. Unitraj: A unified framework for scalable vehi- cle trajectory prediction. arXiv preprint arXiv:2403.15098,

    arXiv

  2. [10]

    Deep inverse reinforcement learning for be- havior prediction in autonomous driving: Accurate forecasts of vehicle motion

    Tharindu Fernando, Simon Denman, Sridha Sridharan, and Clinton Fookes. Deep inverse reinforcement learning for be- havior prediction in autonomous driving: Accurate forecasts of vehicle motion. IEEE Signal Processing Magazine, 38(1): 87–96, 2020. 3

    2020

  3. [11]

    Vectornet: Encoding hd maps and agent dynamics from vectorized rep- resentation

    Jiyang Gao, Chen Sun, Hang Zhao, Yi Shen, Dragomir Anguelov, Congcong Li, and Cordelia Schmid. Vectornet: Encoding hd maps and agent dynamics from vectorized rep- resentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11525– 11...

    2020

  4. [12]

    Thomas: Trajectory heatmap output with learned multi-agent sampling

    Thomas Gilles, Stefano Sabatini, Dzmitry Tsishkou, Bog- dan Stanciulescu, and Fabien Moutarde. Thomas: Trajectory heatmap output with learned multi-agent sampling. arXiv preprint arXiv:2110.06607, 2021. 6

    2021 arXiv

  5. [13]

    Mamba: Linear-time sequence modeling with selective state spaces

    Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 2

    2023 arXiv

  6. [14]

    Densetnt: End-to-end trajectory prediction from dense goal sets

    Junru Gu, Chen Sun, and Hang Zhao. Densetnt: End-to-end trajectory prediction from dense goal sets. In Proceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 15303–15312, 2021. 1, 6, 7

    2021

  7. [15]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025. 2

    2025 arXiv

  8. [16]

    Con- ditional predictive behavior planning with inverse reinforce- ment learning for human-like autonomous driving

    Zhiyu Huang, Haochen Liu, Jingda Wu, and Chen Lv. Con- ditional predictive behavior planning with inverse reinforce- ment learning for human-like autonomous driving. IEEE Transactions on Intelligent Transportation Systems, 2023. 3

    2023

  9. [17]

    Openai o1 system card

    Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richard- son, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, et al. Openai o1 system card. arXiv preprint arXiv:2412.16720, 2024. 2

    2024 arXiv

  10. [18]

    Learning lane graph representa- tions for motion forecasting

    Ming Liang, Bin Yang, Rui Hu, Yun Chen, Renjie Liao, Song Feng, and Raquel Urtasun. Learning lane graph representa- tions for motion forecasting. In European Conference on Computer Vision, pages 541–556, 2020. 2, 4, 5

    2020

  11. [19]

    Kemp: Keyframe-based hierarchical end-to-end deep model for long-term trajectory prediction

    Qiujing Lu, Weiqiao Han, Jeffrey Ling, Minfa Wang, Haoyu Chen, Balakrishnan Varadarajan, and Paul Coving- ton. Kemp: Keyframe-based hierarchical end-to-end deep model for long-term trajectory prediction. In 2022 Inter- national Conference on Robotics and Automation (ICRA) , pa...

    2022

  12. [20]

    Deep learning-based vehicle behavior prediction for autonomous driving applica- tions: A review

    Sajjad Mozaffari, Omar Y Al-Jarrah, Mehrdad Dianati, Paul Jennings, and Alexandros Mouzakitis. Deep learning-based vehicle behavior prediction for autonomous driving applica- tions: A review. IEEE Transactions on Intelligent Trans- portation Systems, 23(1):33–47, 2020. 2

    2020

  13. [21]

    Wayformer: Motion forecasting via simple & efficient attention networks

    Nigamaa Nayakanti, Rami Al-Rfou, Aurick Zhou, Kratarth Goel, et al. Wayformer: Motion forecasting via simple & efficient attention networks. In 2023 International Confer- ence on Robotics and Automation (ICRA), pages 2980–2987. IEEE, 2023. 6, 7

    2023

  14. [22]

    Algorithms for inverse reinforcement learning

    Andrew Y Ng and Stuart Russell. Algorithms for inverse reinforcement learning. In International conference on ma- chine learning, page 2, 2000. 3

    2000

  15. [23]

    Scene transformer: A unified architecture for predicting mul- tiple agent trajectories

    Jiquan Ngiam, Benjamin Caine, Vijay Vasudevan, et al. Scene transformer: A unified architecture for predicting mul- tiple agent trajectories. arXiv preprint arXiv:2106.08417 ,

    arXiv

  16. [24]

    Quadruped robot locomotion in unknown terrain using deep reinforcement learning

    Muleilan Pei, Dongping Wu, and Changhong Wang. Quadruped robot locomotion in unknown terrain using deep reinforcement learning. In 2020 3rd International Confer- ence on Unmanned Systems (ICUS) , pages 517–522. IEEE,

    2020

  17. [25]

    An improved dyna-q algorithm for mobile robot path planning 9 in unknown dynamic environment

    Muleilan Pei, Hao An, Bo Liu, and Changhong Wang. An improved dyna-q algorithm for mobile robot path planning 9 in unknown dynamic environment. IEEE Transactions on Systems, Man, and Cybernetics: Systems , 52(7):4415–4425,

  18. [26]

    Sept: Standard-definition map enhanced scene perception and topology reasoning for au- tonomous driving

    Muleilan Pei, Jiayao Shan, Peiliang Li, Jieqi Shi, Jing Huo, Yang Gao, and Shaojie Shen. Sept: Standard-definition map enhanced scene perception and topology reasoning for au- tonomous driving. IEEE Robotics and Automation Letters , 10(7):7126–7133, 2025. 2

    2025

  19. [27]

    Goirl: Graph-oriented inverse reinforcement learning for multimodal trajectory prediction

    Muleilan Pei, Shaoshuai Shi, Lu Zhang, Peiliang Li, and Shaojie Shen. Goirl: Graph-oriented inverse reinforcement learning for multimodal trajectory prediction. International Conference on Machine Learning (ICML), 2025. 4

    2025

  20. [28]

    Pointnet: Deep learning on point sets for 3d classification and segmentation

    Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652–660,

  21. [29]

    Scene compliant trajectory forecast with agent- centric spatio-temporal grids

    Daniela Ridel, Nachiket Deo, Denis Wolf, and Mohan Trivedi. Scene compliant trajectory forecast with agent- centric spatio-temporal grids. IEEE Robotics and Automa- tion Letters, 5(2):2816–2823, 2020. 2

    2020

  22. [30]

    U- net: Convolutional networks for biomedical image segmen- tation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U- net: Convolutional networks for biomedical image segmen- tation. In International Conference on Medical image com- puting and computer-assisted intervention , pages 234–241. Springer, 2015. 5

    2015

  23. [31]

    Learning agents for uncertain environments

    Stuart Russell. Learning agents for uncertain environments. In Proceedings of the eleventh annual conference on Com- putational learning theory, pages 101–103, 1998. 3

    1998

  24. [32]

    Motion transformer with global intention localization and lo- cal movement refinement

    Shaoshuai Shi, Li Jiang, Dengxin Dai, and Bernt Schiele. Motion transformer with global intention localization and lo- cal movement refinement. Advances in Neural Information Processing Systems, 35:6531–6543, 2022. 1, 3

    2022

  25. [33]

    Pip: Planning- informed trajectory prediction for autonomous driving

    Haoran Song, Wenchao Ding, Yuxuan Chen, Shaojie Shen, Michael Yu Wang, and Qifeng Chen. Pip: Planning- informed trajectory prediction for autonomous driving. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXI 16, page...

    2020

  26. [34]

    Learning to predict vehicle trajectories with model-based planning

    Haoran Song, Di Luan, Wenchao Ding, Michael Y Wang, and Qifeng Chen. Learning to predict vehicle trajectories with model-based planning. In Conference on Robot Learn- ing, pages 1035–1045, 2021. 1, 2

    2021

  27. [35]

    Reinforcement learning: An introduction

    Richard S Sutton, Andrew G Barto, et al. Reinforcement learning: An introduction. MIT press Cambridge, 1998. 2

    1998

  28. [36]

    Hpnet: Dynamic trajectory fore- casting with historical prediction attention

    Xiaolong Tang, Meina Kan, Shiguang Shan, Zhilong Ji, Jin- feng Bai, and Xilin Chen. Hpnet: Dynamic trajectory fore- casting with historical prediction attention. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15261–15270, 2024. 6, 7

    2024

  29. [37]

    Multi- path++: Efficient information fusion and trajectory aggrega- tion for behavior prediction

    Balakrishnan Varadarajan, Ahmed Hefny, Avikalp Srivas- tava, Khaled S Refaat, Nigamaa Nayakanti, et al. Multi- path++: Efficient information fusion and trajectory aggrega- tion for behavior prediction. In 2022 International Confer- ence on Robotics and Automation (ICRA), pages...

    2022

  30. [38]

    Argoverse 2: Next generation datasets for self-driving perception and forecasting

    Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes, et al. Argoverse 2: Next generation datasets for self-driving perception and forecasting. arXiv preprint arXiv:230...

    2023 arXiv

  31. [39]

    Efficient sampling-based maximum entropy inverse reinforcement learning with application to autonomous driving

    Zheng Wu, Liting Sun, Wei Zhan, Chenyu Yang, and Masayoshi Tomizuka. Efficient sampling-based maximum entropy inverse reinforcement learning with application to autonomous driving. IEEE Robotics and Automation Letters, 5(4):5355–5362, 2020. 3

    2020

  32. [40]

    Large-scale cost function learning for path planning using deep inverse reinforcement learning

    Markus Wulfmeier, Dushyant Rao, Dominic Zeng Wang, Pe- ter Ondruska, and Ingmar Posner. Large-scale cost function learning for path planning using deep inverse reinforcement learning. The International Journal of Robotics Research, 36 (10):1073–1087, 2017. 4

    2017

  33. [41]

    Goal- lbp: Goal-based local behavior guided trajectory prediction for autonomous driving

    Zhen Yao, Xin Li, Bo Lang, and Mooi Choo Chuah. Goal- lbp: Goal-based local behavior guided trajectory prediction for autonomous driving. IEEE Transactions on Intelligent Transportation Systems, 2023. 6

    2023

  34. [42]

    Decoupling motion forecasting into directional intentions and dynamic states

    Bozhou Zhang, Nan Song, and Li Zhang. Decoupling motion forecasting into directional intentions and dynamic states. arXiv preprint arXiv:2410.05982, 2024. 6, 7

    2024 arXiv

  35. [43]

    Tra- jectory prediction with graph-based dual-scale context fu- sion

    Lu Zhang, Peiliang Li, Jing Chen, and Shaojie Shen. Tra- jectory prediction with graph-based dual-scale context fu- sion. In 2022 IEEE/RSJ International Conference on Intelli- gent Robots and Systems (IROS), pages 11374–11381. IEEE,

    2022

  36. [44]

    Simpl: A simple and efficient multi-agent motion prediction base- line for autonomous driving

    Lu Zhang, Peiliang Li, Sikang Liu, and Shaojie Shen. Simpl: A simple and efficient multi-agent motion prediction base- line for autonomous driving. IEEE Robotics and Automation Letters, 2024. 5, 6

    2024

  37. [45]

    Hivt: Hierarchical vector transformer for multi-agent motion prediction

    Zikang Zhou, Luyao Ye, Jianping Wang, Kui Wu, and Ke- jie Lu. Hivt: Hierarchical vector transformer for multi-agent motion prediction. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 8823–8833, 2022. 6

    2022

  38. [46]

    Query-centric trajectory prediction

    Zikang Zhou, Jianping Wang, Yung-Hui Li, and Yu-Kai Huang. Query-centric trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 17863–17873, 2023. 1, 3, 6, 7

    2023

  39. [47]

    Vision mamba: Efficient visual representation learning with bidirectional state space model

    Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 2

    2024 arXiv

  40. [48]

    Maximum entropy inverse reinforcement learning

    Brian D Ziebart, Andrew L Maas, J Andrew Bagnell, Anind K Dey, et al. Maximum entropy inverse reinforcement learning. In AAAI, pages 1433–1438. Chicago, IL, USA,

Pith tools

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