Pith. sign in

REVIEW 4 major objections 4 minor 56 references

GoIRL: Graph-Oriented Inverse Reinforcement Learning for Multimodal Trajectory Prediction

T0 review · 4 major / 4 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 maximum-entropy inverse reinforcement learning predictor with graph-based map features matches or beats supervised baselines on Argoverse and nuScenes, and reroutes its plans when the drivable area changes.

desk verdict GoIRL has a real feature-adaptor contribution and a coherent IRL-based predictor, but the SOTA claim on Argoverse is contradicted by its own appendix and the generalization story is anecdotal. read the letter →

arxiv 2506.21121 v1 pith:NNNOCNFU submitted 2025-06-26 cs.CV cs.RO

Muleilan Pei , Shaoshuai Shi , Lu Zhang , Peiliang Li , Shaojie Shen This is my paper
classification cs.CVcs.RO
keywords inversereinforcementlearningmultimodaltrajectorypredictionmaximumentropyIRLgraphneuralnetworksautonomousdrivingmotionforecastingfeatureadaptorMCMCsampling
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

GoIRL claims that inverse reinforcement learning can be made competitive with, and in places better than, today's supervised trajectory predictors. The authors combine maximum-entropy IRL with vectorized graph-based map features through a feature adaptor, then turn the learned reward into multimodal future paths via MCMC plan sampling and a two-scale decoder. They report numbers that beat the comparison methods on the Argoverse and nuScenes motion forecasting benchmarks. They further claim that because the predictor is reward-driven it re-plans around map changes, such as a newly undrivable intersection, without any retraining, which behavior-cloning models cannot do.

What carries the argument

The feature adaptor is the load-bearing component: it maps vectorized lane and drivable-area features into a uniform 50x50 grid aligned with the drivable-area graph, zero-pads undrivable cells, and downsamples to a 25x25 MDP state space. This is what lets a maximum-entropy IRL pipeline, which needs grid-structured states and rewards, consume the graph representations that power modern supervised predictors. Without it, IRL-based forecasters are stuck with rasterized bird's-eye-view inputs; with it, the reward network can combine topology, lane connectivity, and drivable-area geometry in one objective.

What would settle it

Sample a large random subset of Argoverse or nuScenes validation scenes, edit a random connected patch of drivable cells to undrivable at inference time, and measure the fraction of GoIRL's top-K sampled plans that still cross the patch. If that fraction does not drop substantially compared with unmodified scenes, or does not drop more than the fraction for a supervised baseline, the claimed IRL generalization advantage is unsupported.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that the MaxEnt IRL paradigm, long limited to rasterized grid worlds, can be grafted onto vectorized context representation and still reach leaderboard-topping accuracy. The reward function is a neural network whose input is a coarse grid assembled by assigning lane-graph and drivable-area node features to cells. Approximate value iteration converts the reward into a MaxEnt policy; MCMC sampling draws plan sequences that are converted to Bezier-curve proposals, clustered into K modes, refined with fine-grained local features, and scored by the product of a classification probability and the MCMC mode frequency. The paper's reported single-model Argoverse brier-minFDE6 is 1.7957 and its nuScenes minADE5 is 0.86, both better than the baselines in the paper's main comparison tables. The claimed generalization advantage is qualitative: when free cells are edited to undrivable, GoIRL's sampled plans avoid the new obstacle, while the supervised DSP baseline does not.

Load-bearing premise

The generalization claim rests on the assumption that a reward network trained on normal maps, when handed a map with formerly free cells marked undrivable, re-optimizes its policy to avoid them as reliably as if it had seen such edited maps in training, a behavior shown in only three hand-picked examples and no aggregate measurement.

Editorial extensions

If this is right

  • IRL-based predictors can consume vectorized lane-graph context via the feature adaptor, closing the representation gap that previously confined them to rasterized bird's-eye-view inputs.
  • Changing the drivable-area map at inference time changes the policy's sampled plans, so novel map configurations can be handled without retraining the network.
  • Fusing classifier probabilities with MCMC-derived mode frequencies improves the Brier score (confidence quality) over classifier-only scoring, as shown in the ablation table.
  • Reported single-model results on Argoverse (brier-minFDE6 = 1.7957) and nuScenes (minADE5 = 0.86) position the IRL predictor competitively with or above the supervised baselines listed in the paper.

Reading between the lines

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

  • If the rerouting behavior holds on a random sample rather than three hand-picked cases, the same reward could be reused by downstream planners to respond to temporary closures and construction zones by editing the drivable-area input alone.
  • The dependence on a 25x25 grid and a fixed action set means the advantage of the IRL stage is most pronounced for map-driven maneuvers; socially complex interactions may lean heavily on the fine-scale refinement stage, whose contribution the ablations show to be substantial.
  • A natural testable extension is to report the fraction of sampled plans that cross newly undrivable cells over a large validation sample, which would convert the qualitative generalization claim into a quantitative one.
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 / 4 minor

Summary. The paper proposes GoIRL, a two-stage trajectory prediction framework. A graph-based context encoder represents lanes, drivable areas, and agents in vectorized form; a feature adaptor projects these features into a grid; and maximum-entropy inverse reinforcement learning (MaxEnt IRL) is used to infer a reward and policy over that grid. Trajectory proposals are then sampled from the learned policy via MCMC, parameterized as Bézier curves, clustered, and refined, with probabilities fused from the classifier and the sampling distribution. Experiments on Argoverse and nuScenes are reported, along with ablations and qualitative demonstrations aimed at showing generalization to drivable-area changes.

Significance. If the reported performance were correct, GoIRL would be a valuable demonstration that an IRL-based predictor with vectorized graph context can be competitive with supervised forecasting models; the feature-adapter idea is a sensible bridge between graph encoders and grid-based IRL, and the hierarchical decoder is coherent. The paper also includes honest failure cases and a component-wise ablation, which is helpful. However, the central 'state-of-the-art' claim on Argoverse is contradicted by the paper's own appendix table, and the nuScenes comparison is too sparse to support the claimed ranking. The generalization claim rests on qualitative examples only. These issues affect the paper's main stated contributions.

major comments (4)
  1. [Abstract, §4.2, Appendix Table 5] The stated 'state-of-the-art' claim on Argoverse is contradicted by the paper's own appendix. Appendix Table 5 reports QCNet with brier-minFDE6 1.693 and SEPT with 1.682, both below GoIRL's 1.695 on the same metric, while the main Tables 1 and 2 omit these methods. GoIRL's own reported value also changes across tables: 1.7957 in Table 1, 1.6947 in Table 2, and 1.695 in Table 5. Since brier-minFDE6 is identified as the official ranking metric, the paper's own numbers do not show GoIRL as the best entry on Argoverse. The main tables, the abstract, and the conclusion need to be reconciled with the full leaderboard; as written, the central empirical claim is internally inconsistent.
  2. [§4.2, Table 3] The claim of 'top-ranked performance' on nuScenes is not supported by Table 3, which compares only P2T, PGP, MacFormer, and Goal-LBP. These are not sufficient to establish a state-of-the-art ranking on the nuScenes leaderboard, and no recent supervised leaderboard entries are included. The comparison should either cover the relevant leaderboard entries or the claim should be softened to 'competitive on the evaluated baselines.'
  3. [§4.3, Appendix C.2] The claimed superiority of IRL over supervised models under drivable-area changes is supported only by qualitative examples: Figure 5 shows one T-junction scenario and Appendix Figure 8 shows a few additional scenes. There is no aggregate evaluation over a set of systematically modified drivable-area scenarios, no protocol for constructing those modifications, and no quantitative comparison with DSP or any other supervised baseline under the same modifications. A concrete quantitative test is needed before claiming superior generalization.
  4. [§3.4, Algorithm 1] The MaxEnt IRL stage is the core novelty, but the paper does not specify how the terminal state distribution Sgoal is parameterized or trained, nor how the number of value-iteration steps N is chosen. The statement that Algorithm 1 yields the MaxEnt policy after N steps is asserted rather than demonstrated. Please provide the objective for the terminal-state network and at least an ablation over N or a convergence diagnostic.
minor comments (4)
  1. [§4.1, Appendix A.1] The grid size is given as 100×100 in the main text and as 50×50 (HG=WG=50) in the appendix; these need to be reconciled.
  2. [Tables 1, 2, and 5] The tables should state explicitly whether results are on the validation or test split and whether each row is a single model or an ensemble; the appendix currently mixes entries from both categories without clearly distinguishing them.
  3. [Eq. (3)] The gradient expression ∇LD = (µD − E[µ])∇R is notationally imprecise; please expand it to show that the gradient is taken with respect to the reward parameters and that µD and E[µ] are visitation-frequency vectors.
  4. [Figure 3] The labels in Figure 3 are difficult to read in the reproduced PDF; increasing the font size of the submodule labels would improve clarity.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation: MaxEnt reward fitting and supervised decoding are compositional; self-citations to DSP/SIMPL are non-load-bearing, though the SOTA claim conflicts with the paper's own Appendix Table 5.

full rationale

The derivation chain is not circular. The reward R is learned by maximizing the log-likelihood of quantized demonstrations (Eq. 3) via soft value iteration (Algorithm 1), and the trajectory decoder is trained separately under Huber regression and Hinge classification losses (Eq. 12) with winner-takes-all selection. Test outputs are therefore not fitted parameters renamed as predictions: Eq. 10 adds a supervised offset to a Bezier proposal, and Eq. 11 normalizes the product of a learned classifier score and the MCMC frequency Pmcmc, both of which are model outputs rather than copies of the training labels. The drivable-area generalization experiment is a forward-pass sensitivity of the CNN reward to zero-padded cells rather than a tautology; it is only qualitative, so it is weak evidence, not circular. The paper does rely on the authors' own DSP and SIMPL for the graph encoder template and the ensemble strategy, but those are published and externally evaluated, and the central MaxEnt IRL plus feature-adaptor derivation does not reduce to them. One non-circular internal inconsistency: Appendix Table 5 lists QCNet (brier-minFDE6 1.693) and SEPT (1.682) versus GoIRL's 1.695, contradicting the abstract's state-of-the-art claim; this is a correctness and support issue rather than a circularity.

Assumptions & free parameters 8 free parameters · 6 assumptions · 0 invented entities

The central result rests on a specific discretization and a nonstandard inferred terminal distribution. The most important free choices are the grid and horizon geometry, the oversampling and mode counts, and the unreported number of value-iteration steps. The reward model itself is learned, so it is not a free parameter in the ledger sense; it is the output of the first training stage.

free parameters (8)
  • Grid cell resolution = 1.0 m
    Chosen by hand to align with drivable-area nodes; affects quantization of demonstrations and the IRL state space (Section 3.3, Appendix A.1).
  • Coarse grid dimension = 100x100, downsampled to 25x25
    Chosen by hand; downsampling reduces IRL cost, but coarser states may lose trajectory detail (Section 4.1).
  • Planning horizon H = 25
    Chosen by hand; fixes the number of states per plan and the MaxEnt trajectory length (Section 4.1).
  • Bezier curve degree n = 5
    Chosen by hand; controls smoothness and expressiveness of trajectory proposals (Section 4.1).
  • Oversample count L = 600
    Chosen by hand; larger L improves cluster stability at compute cost (Section 4.1).
  • Mode count K = 6 for Argoverse, 10 for nuScenes
    Tied to evaluation protocols (minADE6, minADE10), but still a design choice affecting fusion probabilities (Section 4.1).
  • Loss weights alpha, beta, gamma = alpha=1, beta=1, gamma=3
    Fixed by hand without sensitivity analysis (Section 3.6, Appendix A.2).
  • Number of value iteration steps N = not reported
    Approximate MaxEnt value iteration in Algorithm 1 depends on an unreported N or stopping criterion; performance of the reward and policy depends on it (Section 3.4).
assumptions (6)
  • standard math Maximum entropy IRL distribution P(tau) = exp(R(tau))/Z correctly models expert trajectory distribution.
    Invoked in Section 3.4, Eq. (2), as the foundation of MaxEnt IRL (Ziebart et al., 2008).
  • domain assumption The finite MDP with a 25x25 coarse grid, 9 discrete actions, deterministic transitions, and horizon H=25 is a sufficient abstraction of driving behavior for trajectory prediction.
    Sections 3.1 and 4.1; any behavior not representable on this grid or within 25 steps cannot be predicted.
  • ad hoc to paper The terminal state distribution can be inferred by neural networks rather than specified a priori.
    Section 3.4 states 'the terminal state distribution is also inferred by the neural networks'; this departs from standard MaxEnt IRL and is not justified.
  • ad hoc to paper Approximate value iteration in Algorithm 1 converges to the MaxEnt policy after N steps.
    Algorithm 1, Section 3.4; N and the stopping criterion are never reported, yet the reward and policy training rely on this convergence.
  • domain assumption A degree-5 Bezier curve can represent all relevant future trajectories, including sharp turns and sudden stops.
    Section 3.5, Eq. (8); the decoder constrains every proposal to a fifth-degree Bezier curve.
  • ad hoc to paper Assigning fused drivable-area features to the nearest grid cell and zero-padding undrivable cells preserves lane and drivable-area topology for the IRL reward.
    Section 3.3 Feature Adaptor and Appendix A.1; this operation is the key bridge between vectorized context and grid-based IRL and is not validated in isolation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GoIRL: Graph-Oriented Inverse Reinforcement Learning for Multimodal Trajectory Prediction." pith.science (2026). https://pith.science/paper/NNNOCNFU

@misc{pith2026250621121,
  author       = {Pith},
  title        = {Pith review of: GoIRL: Graph-Oriented Inverse Reinforcement Learning for Multimodal Trajectory Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NNNOCNFU}},
  note         = {Machine review of arXiv:2506.21121}
}
read the original abstract

Trajectory prediction for surrounding agents is a challenging task in autonomous driving due to its inherent uncertainty and underlying multimodality. Unlike prevailing data-driven methods that primarily rely on supervised learning, in this paper, we introduce a novel Graph-oriented Inverse Reinforcement Learning (GoIRL) framework, which is an IRL-based predictor equipped with vectorized context representations. We develop a feature adaptor to effectively aggregate lane-graph features into grid space, enabling seamless integration with the maximum entropy IRL paradigm to infer the reward distribution and obtain the policy that can be sampled to induce multiple plausible plans. Furthermore, conditioned on the sampled plans, we implement a hierarchical parameterized trajectory generator with a refinement module to enhance prediction accuracy and a probability fusion strategy to boost prediction confidence. Extensive experimental results showcase our approach not only achieves state-of-the-art performance on the large-scale Argoverse & nuScenes motion forecasting benchmarks but also exhibits superior generalization abilities compared to existing supervised models.

Figures

Figures reproduced from arXiv: 2506.21121 by the authors.

Figure 1
Figure 1. A motivating example of the covariate shift issue in tra￾jectory prediction. In a T-junction driving scenario, the target agent has two potential future trajectories: going straight and turning left. During the data collection process, the ground-truth trajectory is labeled as going straight. However, during testing, the presence of a tentative barricade set up in the intersection renders the original prediction unr… view at source ↗
Figure 2
Figure 2. Overview of GoIRL, illustrating the generation of multimodal trajectory predictions for the target agent (depicted in blue). then marginalizing over them: P(Y|C ˆ ) = X τˆ∈S(C) P(Y| ˆ τ, ˆ C)P(ˆτ |C), (1) where S(C) denotes the space of plausible plans based on the driving context, which is derived from the policy inference. 3.2. Framework Overview An overview of our proposed framework is presented in Fig￾ure 2, whi… view at source ↗
Figure 3
Figure 3. Structure of Bezier curve-based trajectory proposal de- ´ coder for each sampled plan. state set Sgoal is unknown in the trajectory prediction task, the terminal state distribution is also inferred by the neural networks. Upon convergence of the reward distribution, we can acquire the optimal MaxEnt policy π ∗ , which enables the generation of multiple plans over the 2-D grid, serving as trajectory generation priors… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Visualizations of GoIRL on the Argoverse validation set. The historical trajectory, ground-truth future trajectory, and multimodal predictions are depicted in red, magenta, and green, respectively. Other traffic participants have been excluded to em￾phasize the predict…
Figure 5
Figure 5. Figure 5: Qualitative comparisons between GoIRL (left) and DSP (right) in handling drivable area changes. The original predictions are depicted in blue, while the updated trajectories are in green. The undrivable area is indicated with a “STOP” sign symbolizing the transformed r…
Figure 6
Figure 6. Figure 6: Information flow of agent-map feature fusion. pling drivable nodes centered around the target agent within a 100-meter range and a resolution of 1 meter. These nodes are denoted Vd ∈ R Nd×2 , where Nd is the number of driv￾able nodes. We use eight-neighbor dilated conn…
Figure 7
Figure 7. Figure 7: Snapshots with multimodal predictions in the demo video. STOP STOP STOP Updated predictions Observations Ground truth Undrivable area STOP [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 8
Figure 8. Figure 8: Examples of generalization to changes in drivable areas. The forecasted trajectories are in green, and the undrivable area is indicated with a “STOP” sign symbolizing the transformed region. qualitative results can be found in the latter half of the sup￾plementary vide…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

56 extracted references · 47 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    R., Deshpande, A., and Motional, H

    Afshar, S., Deo, N., Bhagat, A., Chakraborty, T., Shao, Y., Buddharaju, B. R., Deshpande, A., and Motional, H. C. Pbp: Path-based trajectory prediction for autonomous driving. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pp.\ 12927--12934. IEEE, 2024

    work page 2024

  3. [3]

    Neural machine translation by jointly learning to align and translate

    Bahdanau, D., Cho, K., and Bengio, Y. Neural machine translation by jointly learning to align and translate. International Conference on Learning Representations (ICLR), 2015

    work page 2015

  4. [4]

    H., Vora, S., Liong, V

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

    work page 2020

  5. [5]

    Argoverse: 3d tracking and forecasting with rich maps

    Chang, M.-F., Lambert, J., Sangkloy, P., Singh, J., Bak, S., Hartnett, A., Wang, D., Carr, P., Lucey, S., Ramanan, D., et al. Argoverse: 3d tracking and forecasting with rich maps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 8748--8757, 2019

    work page 2019

  6. [6]

    End-to-end autonomous driving: Challenges and frontiers

    Chen, L., Wu, P., Chitta, K., Jaeger, B., Geiger, A., and Li, H. End-to-end autonomous driving: Challenges and frontiers. arXiv, 2306.16927, 2023

    arXiv 2023

  7. [7]

    Multimodal trajectory predictions for autonomous driving using deep convolutional networks

    Cui, H., Radosavljevic, V., Chou, F.-C., Lin, T.-H., Nguyen, T., Huang, T.-K., Schneider, J., and Djuric, N. Multimodal trajectory predictions for autonomous driving using deep convolutional networks. In 2019 international conference on robotics and automation (icra), pp.\ 2090--2096. IEEE, 2019

    work page 2019

  8. [8]

    and Trivedi, M

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

    arXiv 2001

Show all 56 references
  1. [9]

    Multimodal trajectory prediction conditioned on lane-graph traversals

    Deo, N., Wolff, E., and Beijbom, O. Multimodal trajectory prediction conditioned on lane-graph traversals. In Conference on Robot Learning, pp.\ 203--212, 2021

    2021

  2. [10]

    Macformer: Map-agent coupled transformer for real-time and robust trajectory prediction

    Feng, C., Zhou, H., Lin, H., Zhang, Z., Xu, Z., Zhang, C., Zhou, B., and Shen, S. Macformer: Map-agent coupled transformer for real-time and robust trajectory prediction. IEEE Robotics and Automation Letters, 2023

    2023

  3. [11]

    A connection between generative adversarial networks, inverse reinforcement learning, and energy-based models

    Finn, C., Christiano, P., Abbeel, P., and Levine, S. A connection between generative adversarial networks, inverse reinforcement learning, and energy-based models. arXiv preprint arXiv:1611.03852, 2016 a

    2016 arXiv

  4. [12]

    Guided cost learning: Deep inverse optimal control via policy optimization

    Finn, C., Levine, S., and Abbeel, P. Guided cost learning: Deep inverse optimal control via policy optimization. In International conference on machine learning, pp.\ 49--58. PMLR, 2016 b

    2016

  5. [13]

    Learning robust rewards with adverserial inverse reinforcement learning

    Fu, J., Luo, K., and Levine, S. Learning robust rewards with adverserial inverse reinforcement learning. In International Conference on Learning Representations, 2018

    2018

  6. [14]

    Vectornet: Encoding hd maps and agent dynamics from vectorized representation

    Gao, J., Sun, C., Zhao, H., Shen, Y., Anguelov, D., Li, C., and Schmid, C. Vectornet: Encoding hd maps and agent dynamics from vectorized representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 11525--11533, 2020

    2020

  7. [15]

    Dynamic scenario representation learning for motion forecasting with heterogeneous graph convolutional recurrent networks

    Gao, X., Jia, X., Li, Y., and Xiong, H. Dynamic scenario representation learning for motion forecasting with heterogeneous graph convolutional recurrent networks. IEEE Robotics and Automation Letters, 8 0 (5): 0 2946--2953, 2023

    2023

  8. [16]

    Home: Heatmap output for future motion estimation

    Gilles, T., Sabatini, S., Tsishkou, D., Stanciulescu, B., and Moutarde, F. Home: Heatmap output for future motion estimation. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pp.\ 500--507. IEEE, 2021

    2021

  9. [17]

    Thomas: Trajectory heatmap output with learned multi-agent sampling

    Gilles, T., Sabatini, S., Tsishkou, D., Stanciulescu, B., and Moutarde, F. Thomas: Trajectory heatmap output with learned multi-agent sampling. International Conference on Learning Representations (ICLR), 2022 a

    2022

  10. [18]

    Gohome: Graph-oriented heatmap output for future motion estimation

    Gilles, T., Sabatini, S., Tsishkou, D., Stanciulescu, B., and Moutarde, F. Gohome: Graph-oriented heatmap output for future motion estimation. In 2022 international conference on robotics and automation (ICRA), pp.\ 9107--9114. IEEE, 2022 b

    2022

  11. [19]

    A., Kahou, S

    Girgis, R., Golemo, F., Codevilla, F., Weiss, M., D'Souza, J. A., Kahou, S. E., Heide, F., and Pal, C. Latent variable sequential set transformers for joint multi-agent motion prediction. International Conference on Learning Representations (ICLR), 2022

    2022

  12. [20]

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

    Gu, J., Sun, C., and Zhao, H. Densetnt: End-to-end trajectory prediction from dense goal sets. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 15303--15312, 2021

    2021

  13. [21]

    End-to-end trajectory distribution prediction based on occupancy grid maps

    Guo, K., Liu, W., and Pan, J. End-to-end trajectory distribution prediction based on occupancy grid maps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 2242--2251, 2022

    2022

  14. [22]

    Social gan: Socially acceptable trajectories with generative adversarial networks

    Gupta, A., Johnson, J., Fei-Fei, L., Savarese, S., and Alahi, A. Social gan: Socially acceptable trajectories with generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 2255--2264, 2018

    2018

  15. [23]

    Hastings, W. K. Monte carlo sampling methods using markov chains and their applications. Biometrika, 57 0 (1): 0 97--109, 1970

    1970

  16. [24]

    and Ermon, S

    Ho, J. and Ermon, S. Generative adversarial imitation learning. Advances in neural information processing systems, 29, 2016

    2016

  17. [25]

    Multimodal trajectory prediction: A survey

    Huang, R., Xue, H., Pagnucco, M., Salim, F., and Song, Y. Multimodal trajectory prediction: A survey. arXiv preprint arXiv:2302.10463, 2023 a

    2023 arXiv

  18. [26]

    Conditional predictive behavior planning with inverse reinforcement learning for human-like autonomous driving

    Huang, Z., Liu, H., Wu, J., and Lv, C. Conditional predictive behavior planning with inverse reinforcement learning for human-like autonomous driving. IEEE Transactions on Intelligent Transportation Systems, 2023 b

    2023

  19. [27]

    Multimodal deep generative models for trajectory prediction: A conditional variational autoencoder approach

    Ivanovic, B., Leung, K., Schmerling, E., and Pavone, M. Multimodal deep generative models for trajectory prediction: A conditional variational autoencoder approach. IEEE Robotics and Automation Letters, 6 0 (2): 0 295--302, 2020

    2020

  20. [28]

    M., Ziebart, B

    Kitani, K. M., Ziebart, B. D., Bagnell, J. A., and Hebert, M. Activity forecasting. In Computer Vision--ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part IV 12, pp.\ 201--214. Springer, 2012

    2012

  21. [29]

    Socially compliant mobile robot navigation via inverse reinforcement learning

    Kretzschmar, H., Spies, M., Sprunk, C., and Burgard, W. Socially compliant mobile robot navigation via inverse reinforcement learning. The International Journal of Robotics Research, 35 0 (11): 0 1289--1307, 2016

    2016

  22. [30]

    Lan, Z., Jiang, Y., Mu, Y., Chen, C., and Li, S. E. Sept: Towards efficient scene representation learning for motion prediction. arXiv preprint arXiv:2309.15289, 2023

    2023 arXiv

  23. [31]

    Learning lane graph representations for motion forecasting

    Liang, M., Yang, B., Hu, R., Chen, Y., Liao, R., Feng, S., and Urtasun, R. Learning lane graph representations for motion forecasting. In European Conference on Computer Vision, pp.\ 541--556. Springer, 2020

    2020

  24. [32]

    Feature pyramid networks for object detection

    Lin, T.-Y., Doll \'a r, P., Girshick, R., He, K., Hariharan, B., and Belongie, S. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 2117--2125, 2017

    2017

  25. [33]

    S., and Sapp, B

    Nayakanti, N., Al-Rfou, R., Zhou, A., Goel, K., Refaat, K. S., and Sapp, B. Wayformer: Motion forecasting via simple & efficient attention networks. arXiv preprint arXiv:2207.05844, 2022

    2022 arXiv

  26. [34]

    Ng, A. Y. and Russell, S. J. Algorithms for inverse reinforcement learning. In Proceedings of the Seventeenth International Conference on Machine Learning, pp.\ 663--670, 2000

    2000

  27. [35]

    A., Abbeel, P., Peters, J., et al

    Osa, T., Pajarinen, J., Neumann, G., Bagnell, J. A., Abbeel, P., Peters, J., et al. An algorithmic perspective on imitation learning. Foundations and Trends in Robotics , 7 0 (1-2): 0 1--179, 2018

    2018

  28. [36]

    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 Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 652--660, 2017

    2017

  29. [37]

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

    Ridel, D., Deo, N., Wolf, D., and Trivedi, M. Scene compliant trajectory forecast with agent-centric spatio-temporal grids. IEEE Robotics and Automation Letters, 5 0 (2): 0 2816--2823, 2020

    2020

  30. [38]

    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

  31. [39]

    Y., and Chen, Q

    Song, H., Luan, D., Ding, W., Wang, M. Y., and Chen, Q. Learning to predict vehicle trajectories with model-based planning. In Conference on Robot Learning, pp.\ 1035--1045, 2021

    2021

  32. [40]

    S., Nayakanti, N., Cornman, A., Chen, K., Douillard, B., Lam, C

    Varadarajan, B., Hefny, A., Srivastava, A., Refaat, K. S., Nayakanti, N., Cornman, A., Chen, K., Douillard, B., Lam, C. P., Anguelov, D., et al. Multipath++: Efficient information fusion and trajectory aggregation for behavior prediction. In 2022 International Conference on Ro...

    2022

  33. [41]

    Attention is all you need

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A., Kaiser, L., and I., P. Attention is all you need. Advances in Neural Information Processing Systems, pp.\ 5998--6008, 2017

    2017

  34. [42]

    Ganet: Goal area network for motion forecasting

    Wang, M., Zhu, X., Yu, C., Li, W., Ma, Y., Jin, R., Ren, X., Ren, D., Wang, M., and Yang, W. Ganet: Goal area network for motion forecasting. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pp.\ 1609--1615. IEEE, 2023

    2023

  35. [43]

    Z., Ondruska, P., and Posner, I

    Wulfmeier, M., Rao, D., Wang, D. Z., Ondruska, P., and Posner, I. Large-scale cost function learning for path planning using deep inverse reinforcement learning. The International Journal of Robotics Research, 36 0 (10): 0 1073--1087, 2017

    2017

  36. [44]

    Improving out-of-distribution generalization of trajectory prediction for autonomous driving via polynomial representations

    Yao, Y., Yan, S., Goehring, D., Burgard, W., and Reichardt, J. Improving out-of-distribution generalization of trajectory prediction for autonomous driving via polynomial representations. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp.\ ...

    2024

  37. [45]

    Yao, Z., Li, X., Lang, B., and Chuah, M. C. Goal-lbp: Goal-based local behavior guided trajectory prediction for autonomous driving. IEEE Transactions on Intelligent Transportation Systems, 2023

    2023

  38. [46]

    Tpcn: Temporal point cloud networks for motion forecasting

    Ye, M., Cao, T., and Chen, Q. Tpcn: Temporal point cloud networks for motion forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 11318--11327, 2021

    2021

  39. [47]

    Meta-inverse reinforcement learning with probabilistic context variables

    Yu, L., Yu, T., Finn, C., and Ermon, S. Meta-inverse reinforcement learning with probabilistic context variables. Advances in neural information processing systems, 32, 2019

    2019

  40. [48]

    End-to-end interpretable neural motion planner

    Zeng, W., Luo, W., Suo, S., Sadat, A., Yang, B., Casas, S., and Urtasun, R. End-to-end interpretable neural motion planner. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 8660--8669, 2019

    2019

  41. [49]

    Lanercnn: Distributed representations for graph-centric motion forecasting

    Zeng, W., Liang, M., Liao, R., and Urtasun, R. Lanercnn: Distributed representations for graph-centric motion forecasting. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp.\ 532--539. IEEE, 2021

    2021

  42. [50]

    Trajectory prediction with graph-based dual-scale context fusion

    Zhang, L., Li, P., Chen, J., and Shen, S. Trajectory prediction with graph-based dual-scale context fusion. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp.\ 11374--11381. IEEE, 2022

    2022

  43. [51]

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

    Zhang, L., Li, P., Liu, S., and Shen, S. Simpl: A simple and efficient multi-agent motion prediction baseline for autonomous driving. IEEE Robotics and Automation Letters, 2024

    2024

  44. [52]

    Integrating kinematics and environment context into deep inverse reinforcement learning for predicting off-road vehicle trajectories

    Zhang, Y., Wang, W., Bonatti, R., Maturana, D., and Scherer, S. Integrating kinematics and environment context into deep inverse reinforcement learning for predicting off-road vehicle trajectories. In Conference on Robot Learning, pp.\ 894--905. PMLR, 2018

    2018

  45. [53]

    Hivt: Hierarchical vector transformer for multi-agent motion prediction

    Zhou, Z., Ye, L., Wang, J., Wu, K., and Lu, K. Hivt: Hierarchical vector transformer for multi-agent motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 8823--8833, 2022

    2022

  46. [54]

    Query-centric trajectory prediction

    Zhou, Z., Wang, J., Li, Y.-H., and Huang, Y.-K. Query-centric trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 17863--17873, 2023

    2023

  47. [55]

    D., Maas, A

    Ziebart, B. D., Maas, A. L., Bagnell, J. A., Dey, A. K., et al. Maximum entropy inverse reinforcement learning. In AAAI, volume 8, pp.\ 1433--1438. Chicago, IL, USA, 2008

    2008

  48. [56]

    D., Ratliff, N., Gallagher, G., Mertz, C., Peterson, K., Bagnell, J

    Ziebart, B. D., Ratliff, N., Gallagher, G., Mertz, C., Peterson, K., Bagnell, J. A., Hebert, M., Dey, A. K., and Srinivasa, S. Planning-based prediction for pedestrians. In 2009 IEEE/RSJ International Conference on Intelligent Robots and Systems, pp.\ 3931--3936. IEEE, 2009

    2009

Pith tools

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