Pith. sign in

REVIEW 3 major objections 5 minor 39 references

Goal-conditioned Hierarchical Reinforcement Learning for Sample-efficient and Safe Autonomous Driving at Intersections

T0 review · 3 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 A hierarchical reinforcement learning agent that masks unsafe subgoals before choosing them reports 94.7% success and 3.3% collisions at simulated intersections.

desk verdict A solid empirical integration of goal-conditioned trajectory prediction into hierarchical RL, but the safety language outruns what the paper actually measures. read the letter →

arxiv 2506.16336 v1 pith:QNICGBPQ submitted 2025-06-19 cs.RO cs.MA

Yiou Huang This is my paper
classification cs.ROcs.MA
keywords hierarchicalreinforcementlearninggoal-conditionedcollisionpredictionintersectionnavigationautonomousdrivingtrajectorysafesampleefficiencySMARTSsimulation
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

The paper aims to show that a two-level hierarchical reinforcement learning agent can learn to drive through intersections with less data and fewer collisions than a single-level RL agent. The key move is to separate what to aim for from how to steer: a high-level decision-maker picks a subgoal from a small set of candidate points, and a low-level motion-planner executes it. A goal-conditioned collision prediction module forecasts, for each candidate subgoal, where the ego vehicle and nearby vehicles will be in the next second, and masks out subgoals whose predicted paths collide. In simulated left-turn, straight, and right-turn intersection tasks, the paper reports a 94.7% overall success rate and a 3.3% collision rate, versus 74.0% success and 10.0% collisions for the best flat-RL baseline tested. This matters because sample efficiency and safety are the two barriers that keep reinforcement learning from being practical for real driving decisions.

What carries the argument

The load-bearing mechanism is the Goal-Conditioned Collision Prediction (GCCP) module, a learned trajectory predictor plus a rule-based collision checker. The predictor takes each candidate subgoal of the ego vehicle, along with the history trajectories, route information, and drivable area of all vehicles, and outputs one-second-ahead trajectories for the ego vehicle and up to five surrounding vehicles; a Separating Axis Theorem overlap test then marks any subgoal whose predicted ego path intersects a predicted surrounding path as forbidden. That collision risk mask is directly added to the decision-maker's logits, forcing the probability of unsafe subgoals to zero. Because the predictions are conditioned on the ego's own candidate subgoals, the mask can reflect how surrounding vehicles would react if the ego yielded, passed, or changed lanes, rather than assuming a single most-likely intention.

What would settle it

Run the trained system in a simulator where, after training, surrounding vehicles brake, accelerate, or cut in at rates outside the training distribution, and check whether any subgoal that survives the collision mask leads to a collision within the one-second prediction horizon; if the masked collision rate rises to the flat-RL baseline, the safety claim fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that intersection navigation can be decomposed into a safe subgoal-selection problem and a subgoal-tracking problem, and that doing so makes reinforcement learning both faster to train and safer at test time. The decision-maker's action space is a fixed set of twelve subgoals sampled from the ego route; the GCCP module predicts the future trajectories of the ego vehicle and surrounding vehicles for each subgoal, converts those predictions into a binary collision risk mask via a geometric overlap check, and adds large penalties to unsafe logits before the decision-maker's softmax. The low-level motion-planner is rewarded for minimizing distance and heading error to the chosen subgoal, with penalties for collision and off-road events. The training procedure deliberately stops updating the motion-planner after episode 500 and only feeds the collision mask to the decision-maker after episode 800, so the trajectory predictor learns against a stable ego-driving policy. The paper reports that this pipeline reaches over 70% success by episode 500 and converges to a steady success rate above 90%, with testing collision rates of 3.3% overall, compared with 10.0% for the strongest flat baseline.

Load-bearing premise

The scheme's safety rests on two linked assumptions: the learned predictor's one-second forecasts of surrounding vehicles are accurate for the subgoals actually chosen, and the low-level motion-planner stays the same after training so those forecasts remain valid.

Editorial extensions

If this is right

  • With the same high-level subgoal set, the motion-planner's skill at approaching a subgoal transfers across different intersection tasks, which is why the paper finds faster convergence than flat RL.
  • The collision mask removes unsafe subgoals before the policy samples, so the decision-maker never needs to explore dangerous actions; the reported 3.3% test collision rate follows directly from that masking.
  • Goal-conditioning the predictor matters: ablating it raises collisions to 4.7%, removing GCCP entirely raises them to 11.3%, and switching to a constant-velocity predictor raises them to 8.7%.
  • Because the mask is computed from predicted trajectories rather than from a learned value function, the safety mechanism stays interpretable: an observer can see which subgoal caused a collision warning.

Reading between the lines

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

  • Editorial extension: the safety guarantee is only as strong as the predictor's forecast; in real traffic with drivers who behave outside the training distribution, the mask could clear a subgoal that later leads to a collision, so the figures should be read as simulator-level rather than absolute guarantees.
  • Editorial extension: nothing in the paper requires the decision-maker to be PPO, so the same subgoal-and-mask structure could likely be retrofitted onto any discrete-action RL head; testing that would isolate whether the gains come from the hierarchy, the mask, or the specific optimizer.
  • Editorial extension: the one-second prediction horizon is matched to the motion-planner's decision cadence; extending the same framework to high-speed roads or multi-lane merges would require longer-horizon prediction or a dynamic horizon.
  • Editorial extension: a direct test of the claimed safety margin would be to deliberately bias surrounding vehicles' behavior after training and see how often the masked set still contains a colliding subgoal; the paper does not report this stress test.
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 / 5 minor

Summary. The paper proposes a hierarchical reinforcement learning (HRL) framework for autonomous driving at intersections, with a high-level decision-maker that selects subgoals and a low-level motion-planner that executes them. A goal-conditioned collision prediction (GCCP) module predicts future trajectories of the ego and surrounding vehicles conditioned on each candidate subgoal, and a rule-based collision checker produces a hard mask that forbids unsafe subgoals. The motion-planner is frozen after 500 episodes, the trajectory predictor is trained online from episode 500 onward, and the safety mask is activated only after episode 800. Experiments in the SMARTS simulator on left-turn, straight, and right-turn intersection tasks compare the method with DQN, PPO, and SAC-Discrete, plus three ablations, reporting a 94.7% overall success rate and 3.3% collision rate versus 74.0% and 10.0% for the best flat RL baseline.

Significance. If substantiated, the contribution is of moderate significance: conditioning collision prediction on the ego vehicle's candidate subgoals is a reasonable and potentially interpretable way to combine hierarchical RL with safety screening, and the paper's ablation structure is a sensible first step. The comparison against three flat RL baselines and the inclusion of a kinematic-predictor ablation are useful. However, the central safety claim is considerably stronger than the evidence, and the headline quantitative results are reported without variance or significance testing. The paper would be strengthened substantially by multi-seed results, confidence intervals, and a closed-loop evaluation of the safety mask's false-negative rate.

major comments (3)
  1. [§III.C and §IV.C.2] The safety claim is contradicted by the paper's own numbers. Equations (1)-(2) implement a hard veto: unsafe subgoals get logit -1e8 and the probability of selecting them is 'effectively zero.' Yet Table I reports a 3.3% collision rate for the full method. After the mask is activated, any collision must be a false negative of the GCCP module, so the safety guarantee is exactly the predictor's correctness. The paper reports only aggregate ADE/FDE in Fig. 7, which does not measure the mask's false-negative rate at decision-relevant moments. Please report, at minimum, the fraction of executed subgoals that were classified safe but led to a collision, and ideally the full confusion matrix of the mask against the rollout outcomes.
  2. [§IV.C, Tables I-II and Figs. 5-6] The headline claims of 'significantly safer' and 'converges faster' are not supported by the reported statistics. The tables and training curves appear to be single point estimates from one run or one fixed test set, with no variance, number of seeds, or significance tests. For example, the difference between the proposed method's 3.3% collision rate and the w/ CV ablation's 8.7% could easily be within run-to-run noise. Please report means and standard deviations over at least five independent training runs, or give confidence intervals and a significance test, and clarify whether the '50 distinct traffic flows' are fixed or resampled across runs.
  3. [§III.E and Algorithm 1] The online training procedure couples the safety mask to the frozen low-level policy and to the pre-mask subgoal distribution in a way that is not evaluated. The motion-planner is frozen at episode 500, the predictor is trained only afterward, and the mask is activated only after episode 800, while the decision-maker continues updating throughout training. This means the predictor is trained on data generated by the decision-maker before the mask is active, and the paper does not check whether the predictor remains calibrated for the final policy's subgoal choices. A closed-loop evaluation of the mask's missed detections and false alarms under the deployed subgoal distribution, as well as under a deployment shift such as denser traffic or unseen intersection geometry, is needed before the paper can claim that safety is 'ensured throughout the decision-making process.'
minor comments (5)
  1. [Abstract and Section II.C] The text contains several typos and grammatical errors, including 'In additional' in the abstract, 'propses' in Section II.C, 'repesented' in Section III.D.1, and 'collsion' in Equation (9). A careful proofreading pass is needed.
  2. [§III.C.2, Eq. (1)] The masking constant is written as '−108'; if this is intended to be -10^8, it should be written unambiguously as -1e8 or -10^8 to avoid confusion with the integer -108.
  3. [§IV.A.2] The description of the 'w/o goal condition' ablation is ambiguous: it says the subgoal feature is replaced with a zero vector when predicting the trajectories of surrounding vehicles, but the architecture in Fig. 2 uses a single subgoal encoder for both ego and surrounding trajectories. Please clarify exactly which inputs are modified and whether the ego trajectory prediction is also affected.
  4. [Figs. 5-6] The training curves would be far more informative with shaded standard deviations or percentile bands across seeds; as presented, the 'converges faster' claim cannot be distinguished from noise.
  5. [Abstract and Conclusions] The claim that subgoal policies can be reused 'across similar tasks for various navigation scenarios' is not directly tested: the experiments cover three intersection tasks that are trained together, but no generalization test to unseen intersection layouts or traffic densities is reported.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the central claims are supported by external baselines, ablations, and closed-loop simulation results.

full rationale

The paper is an empirical systems paper, not a derivation, and its main performance claims are benchmarked against external flat RL baselines (DQN, PPO, SAC-Discrete) and against ablated versions of the framework (w/o GCCP, w/o goal condition, w/ CV) in the SMARTS simulator. The reported success and collision rates are measured in closed-loop testing, not obtained by construction from the training objective. The GCCP trajectory predictor is a learned component trained on replay data, and its contribution is evaluated by ablation rather than asserted from its own training loss. Equations (1) and (2) show that the collision risk mask is a hard veto applied to decision-maker logits, so the safety guarantee is conditional on the predictor's accuracy; this is a correctness and robustness concern, not circularity. The paper explicitly acknowledges the coupling between the predictor and the frozen motion-planner in Section III.E, stating that the motion-planner is stopped at episode Ng=500 and the mask is only activated at episode Nm=800 to let the predictor learn a stable ego trajectory. That is a stated limitation about distribution shift and predictor calibration, not a circular step. There is no load-bearing self-citation chain, no imported uniqueness theorem, and no fitted parameter that is later renamed as a prediction. The central claims therefore have independent empirical grounding, and no circular step meeting the required evidentiary standard can be identified.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim rests on hand-chosen reward weights, subgoal and action discretizations, and a training schedule that freezes the low-level policy. No new physical entities are introduced.

free parameters (4)
  • Reward weights for decision-maker and motion-planner = r_goal=3, r_subgoal=0.5, r_near=-0.5, eps_s=0.05, r_time=-0.05, r_arr=1.0, r_col=-1, r_off=-1, eps_d=0.05, eps_h=0.5
    Chosen by hand in Sections III.D.5 and III.D.9 to shape the learned behavior; no sensitivity analysis is provided.
  • Subgoal sampling density = N=12 subgoals, 5-meter spacing within 20 meters
    Section III.B.2; the size of the decision-maker's action space and the granularity of safety evaluation depend on this hand-chosen value.
  • Motion-planner action set = 6 discrete actions (slow down, keep direction, slowly turn left/right, quickly turn left/right)
    Section III.D.7; the low-level policy's expressiveness is fixed by this hand-designed set.
  • Training schedule thresholds = Ng=500 (freeze motion-planner), Nm=800 (activate collision mask)
    Section III.E; these are hand-chosen and load-bearing because the trajectory predictor only produces a stable ego-trajectory estimate after the motion-planner is frozen.
assumptions (4)
  • domain assumption The intersection navigation problem is a Markov Decision Process with fully observed state including surrounding vehicles' planned routes and task goals.
    Section III.A and III.B.1. In real traffic, other drivers' routes are not observable; this assumption is needed for the state encoder.
  • domain assumption The goal-conditioned trajectory predictor's forecasts are accurate enough for the collision checker.
    Section III.C. The safety mask is only as good as the predicted trajectories; the paper does not quantify the safety-critical error rate.
  • ad hoc to paper The low-level motion-planner's policy becomes stable after 500 episodes, so the ego trajectory can be predicted from the subgoal.
    Section III.E explicitly stops updating the motion-planner after Ng=500 to keep the ego trajectory predictable; this is a design choice specific to this framework.
  • domain assumption The SMARTS simulator provides valid ground-truth dynamics for evaluating success and collision.
    All results are measured in simulation; the paper does not address sim-to-real transfer.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Goal-conditioned Hierarchical Reinforcement Learning for Sample-efficient and Safe Autonomous Driving at Intersections." pith.science (2026). https://pith.science/paper/QNICGBPQ

@misc{pith2026250616336,
  author       = {Pith},
  title        = {Pith review of: Goal-conditioned Hierarchical Reinforcement Learning for Sample-efficient and Safe Autonomous Driving at Intersections},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QNICGBPQ}},
  note         = {Machine review of arXiv:2506.16336}
}
read the original abstract

Reinforcement learning (RL) exhibits remarkable potential in addressing autonomous driving tasks. However, it is difficult to train a sample-efficient and safe policy in complex scenarios. In this article, we propose a novel hierarchical reinforcement learning (HRL) framework with a goal-conditioned collision prediction (GCCP) module. In the hierarchical structure, the GCCP module predicts collision risks according to different potential subgoals of the ego vehicle. A high-level decision-maker choose the best safe subgoal. A low-level motion-planner interacts with the environment according to the subgoal. Compared to traditional RL methods, our algorithm is more sample-efficient, since its hierarchical structure allows reusing the policies of subgoals across similar tasks for various navigation scenarios. In additional, the GCCP module's ability to predict both the ego vehicle's and surrounding vehicles' future actions according to different subgoals, ensures the safety of the ego vehicle throughout the decision-making process. Experimental results demonstrate that the proposed method converges to an optimal policy faster and achieves higher safety than traditional RL methods.

Figures

Figures reproduced from arXiv: 2506.16336 by the authors.

Figure 1
Figure 1. An overview of the framework. The GCCP module outputs the [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. An overview of the goal-conditioned trajectory predictor: The [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The design of the hierarchical decision-making and motion [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Three designed intersection scenarios in the SMARTS simulator [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 6
Figure 6. Figure 6: The training results of the proposed framework and ablated methods [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: The trajectory prediction accuracy of the proposed framework and [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: Three scenarios of the goal-conditioned prediction. The red [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 18 canonical work pages

  1. [1]

    Autonomous Driving at Unsignalized Intersections: A Review of Decision-Making Challenges and Reinforcement Learning-Based Solutions

    M. Al-Sharman, L. Edes, B. Sun, V . Jayakumar, M. A. Daoud, D. Rayside, and W. Melek, “Autonomous driving at unsignalized intersections: A review of decision-making challenges and reinforcement learning-based solutions,” 2024. [Online]. Available: https://arxiv.org/abs/2409.13144

    work page Pith review arXiv 2024

  2. [2]

    Enhanced intelligent driver model to access the impact of driving strategies on traffic capacity,

    A. Kesting, M. Treiber, and D. Helbing, “Enhanced intelligent driver model to access the impact of driving strategies on traffic capacity,” Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences , vol. 368, no. 1928, p. 4585–4605, Oct. 2010. [Online]. Available: http: //dx.doi.org/10.1098/rsta.2010.0084

    work page arXiv 1928

  3. [3]

    Extended safety descriptor measurements for relative safety assessment in mixed road traffic,

    O. Derbel, B. Mourllion, and M. Basset, “Extended safety descriptor measurements for relative safety assessment in mixed road traffic,” in 2012 15th International IEEE Conference on Intelligent Transporta- tion Systems, 2012, pp. 752–757

    work page 2012

  4. [4]

    Analysis of optimal velocity model with explicit delay,

    M. Bando, K. Hasebe, K. Nakanishi, and A. Nakayama, “Analysis of optimal velocity model with explicit delay,” Physical Review E, vol. 58, no. 5, p. 5429–5435, Nov. 1998. [Online]. Available: http://dx.doi.org/10.1103/PhysRevE.58.5429

    doi:10.1103/physreve.58.5429 1998

  5. [5]

    End to end learning for self-driving cars,

    M. Bojarski, D. D. Testa, D. Dworakowski, B. Firner, B. Flepp, P. Goyal, L. D. Jackel, M. Monfort, U. Muller, J. Zhang, X. Zhang, J. Zhao, and K. Zieba, “End to end learning for self-driving cars,”

    work page

  6. [6]

    Chauffeurnet: Learning to drive by imitating the best and synthesizing the worst,

    M. Bansal, A. Krizhevsky, and A. Ogale, “Chauffeurnet: Learning to drive by imitating the best and synthesizing the worst,” 2018. [Online]. Available: https://arxiv.org/abs/1812.03079

    arXiv 2018

  7. [7]

    Navigating Occluded Intersections with Autonomous Vehicles using Deep Reinforcement Learning

    D. Isele, R. Rahimi, A. Cosgun, K. Subramanian, and K. Fujimura, “Navigating occluded intersections with autonomous vehicles using deep reinforcement learning,” 2018. [Online]. Available: https: //arxiv.org/abs/1705.01196

    work page Pith review arXiv 2018

  8. [8]

    Learning Negotiating Behavior Between Cars in Intersections using Deep Q-Learning

    T. Tram, A. Jansson, R. Gr ¨onberg, M. Ali, and J. Sj ¨oberg, “Learning negotiating behavior between cars in intersections using deep q- learning,” 2018. [Online]. Available: https://arxiv.org/abs/1810.10469

    work page Pith review arXiv 2018

Show all 39 references
  1. [9]

    Social attention for autonomous decision-making in dense traffic,

    E. Leurent and J. Mercat, “Social attention for autonomous decision-making in dense traffic,” 2019. [Online]. Available: https: //arxiv.org/abs/1911.12250

    2019 arXiv

  2. [10]

    A multi-task reinforcement learning approach for navigating unsignalized intersections,

    S. Kai, B. Wang, D. Chen, J. Hao, H. Zhang, and W. Liu, “A multi-task reinforcement learning approach for navigating unsignalized intersections,” in 2020 IEEE Intelligent Vehicles Symposium (IV) , 2020, pp. 1583–1588

    2020

  3. [11]

    Pomdp and hierarchical options mdp with continuous actions for autonomous driving at intersections,

    Z. Qiao, K. Muelling, J. Dolan, P. Palanisamy, and P. Mudalige, “Pomdp and hierarchical options mdp with continuous actions for autonomous driving at intersections,” in 2018 21st International Con- ference on Intelligent Transportation Systems (ITSC), 2018, pp. 2377– 2382

    2018

  4. [12]

    Behavior planning at urban intersections through hierarchical reinforcement learning,

    Z. Qiao, J. Schneider, and J. M. Dolan, “Behavior planning at urban intersections through hierarchical reinforcement learning,” 2020. [Online]. Available: https://arxiv.org/abs/2011.04697

    2020 arXiv

  5. [13]

    Action and trajectory planning for urban autonomous driving with hierarchical reinforcement learning,

    X. Lu, F. X. Fan, and T. Wang, “Action and trajectory planning for urban autonomous driving with hierarchical reinforcement learning,”

  6. [14]

    Safe reinforcement learning on autonomous vehicles,

    D. Isele, A. Nakhaei, and K. Fujimura, “Safe reinforcement learning on autonomous vehicles,” in 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , 2018, pp. 1–6

    2018

  7. [15]

    Learning to navigate intersections with unsupervised driver trait inference,

    S. Liu, P. Chang, H. Chen, N. Chakraborty, and K. Driggs-Campbell, “Learning to navigate intersections with unsupervised driver trait inference,” 2022. [Online]. Available: https://arxiv.org/abs/2109.06783

    2022 arXiv

  8. [16]

    Human-like decision making at unsignalized intersections using social value ori- entation,

    Y . Tong, L. Wen, P. Cai, D. Fu, S. Mao, B. Shi, and Y . Li, “Human-like decision making at unsignalized intersections using social value ori- entation,” IEEE Intelligent Transportation Systems Magazine , vol. 16, no. 2, pp. 55–69, 2024

    2024

  9. [17]

    Interactive autonomous navigation with internal state inference and interactivity estimation,

    J. Li, D. Isele, K. Lee, J. Park, K. Fujimura, and M. J. Kochenderfer, “Interactive autonomous navigation with internal state inference and interactivity estimation,” IEEE Transactions on Robotics , vol. 40, pp. 2932–2949, 2024

    2024

  10. [18]

    Between mdps and semi-mdps: A framework for temporal abstraction in reinforcement learning,

    R. S. Sutton, D. Precup, and S. Singh, “Between mdps and semi-mdps: A framework for temporal abstraction in reinforcement learning,” Artificial Intelligence , vol. 112, no. 1, pp. 181–211, 1999. [Online]. Available: https://www.sciencedirect.com/science/article/pii/ S0004370299000521

    1999

  11. [19]

    Feudal reinforcement learning,

    P. Dayan and G. E. Hinton, “Feudal reinforcement learning,” in Advances in Neural Information Processing Systems , S. Hanson, J. Cowan, and C. Giles, Eds., vol. 5. Morgan-Kaufmann, 1992. [Online]. Available: https://proceedings.neurips.cc/paper files/paper/ 1992/file/d14220ee6...

    1992

  12. [20]

    Hierarchical deep reinforcement learning: Integrating temporal abstraction and intrinsic motivation,

    T. D. Kulkarni, K. R. Narasimhan, A. Saeedi, and J. B. Tenenbaum, “Hierarchical deep reinforcement learning: Integrating temporal abstraction and intrinsic motivation,” 2016. [Online]. Available: https://arxiv.org/abs/1604.06057

    2016 arXiv

  13. [21]

    Data-efficient hierarchical reinforcement learning,

    O. Nachum, S. Gu, H. Lee, and S. Levine, “Data-efficient hierarchical reinforcement learning,” 2018. [Online]. Available: https://arxiv.org/abs/1805.08296

    2018 arXiv

  14. [22]

    Learning multi-level hierarchies with hindsight,

    A. Levy, G. Konidaris, R. Platt, and K. Saenko, “Learning multi-level hierarchies with hindsight,” 2019. [Online]. Available: https://arxiv.org/abs/1712.00948

    2019 arXiv

  15. [23]

    Multipath: Multiple probabilistic anchor trajectory hypotheses for behavior prediction,

    Y . Chai, B. Sapp, M. Bansal, and D. Anguelov, “Multipath: Multiple probabilistic anchor trajectory hypotheses for behavior prediction,”

  16. [24]

    Learning lane graph representations for motion forecasting,

    M. Liang, B. Yang, R. Hu, Y . Chen, R. Liao, S. Feng, and R. Urtasun, “Learning lane graph representations for motion forecasting,” 2020. [Online]. Available: https://arxiv.org/abs/2007.13732

    2020 arXiv

  17. [25]

    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,” 2020. [Online]. Available: https://arxiv.org/abs/2008.08294

    2020 arXiv

  18. [26]

    H. Song, W. Ding, Y . Chen, S. Shen, M. Y . Wang, and Q. Chen, PiP: Planning-Informed Trajectory Prediction for Autonomous Driving . Springer International Publishing, 2020, p. 598–614. [Online]. Available: http://dx.doi.org/10.1007/978-3-030-58589-1 36

    2020 doi

  19. [27]

    Learning interaction- aware motion prediction model for decision-making in autonomous driving,

    Z. Huang, H. Liu, J. Wu, W. Huang, and C. Lv, “Learning interaction- aware motion prediction model for decision-making in autonomous driving,” 2023. [Online]. Available: https://arxiv.org/abs/2302.03939

    2023 arXiv

  20. [28]

    Scene transformer: A unified architecture for predicting multiple agent trajectories,

    J. Ngiam, B. Caine, V . Vasudevan, Z. Zhang, H.-T. L. Chiang, J. Ling, R. Roelofs, A. Bewley, C. Liu, A. Venugopal, D. Weiss, B. Sapp, Z. Chen, and J. Shlens, “Scene transformer: A unified architecture for predicting multiple agent trajectories,” 2022. [Online]. Available: htt...

    2022 arXiv

  21. [29]

    Vectornet: Encoding HD maps and agent dynamics from vectorized representation,

    J. Gao, C. Sun, H. Zhao, Y . Shen, D. Anguelov, C. Li, and C. Schmid, “Vectornet: Encoding HD maps and agent dynamics from vectorized representation,” CoRR, vol. abs/2005.04259, 2020. [Online]. Available: https://arxiv.org/abs/2005.04259

    2005 arXiv

  22. [30]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” CoRR, vol. abs/1706.03762, 2017. [Online]. Available: http://arxiv.org/abs/1706.03762

    2017 arXiv

  23. [31]

    Separating axis theorem for oriented bounding boxes,

    J. Huynh, “Separating axis theorem for oriented bounding boxes,”

  24. [32]

    Proximal policy optimization algorithms,

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Proximal policy optimization algorithms,” CoRR, vol. abs/1707.06347, 2017. [Online]. Available: http://arxiv.org/abs/ 1707.06347

    2017 arXiv

  25. [33]

    Smarts: Scalable multi-agent reinforcement learning training school for autonomous driving,

    M. Zhou, J. Luo, J. Villella, Y . Yang, D. Rusu, J. Miao, W. Zhang, M. Alban, I. Fadakar, Z. Chen, A. C. Huang, Y . Wen, K. Hassanzadeh, D. Graves, D. Chen, Z. Zhu, N. Nguyen, M. Elsayed, K. Shao, S. Ahilan, B. Zhang, J. Wu, Z. Fu, K. Rezaee, P. Yadmellat, M. Rohani, N. P. Nie...

    2020 arXiv

  26. [34]

    Playing atari with deep reinforcement learning,

    V . Mnih, K. Kavukcuoglu, D. Silver, A. Graves, I. Antonoglou, D. Wierstra, and M. A. Riedmiller, “Playing atari with deep reinforcement learning,” CoRR, vol. abs/1312.5602, 2013. [Online]. Available: http://arxiv.org/abs/1312.5602

    2013 arXiv

  27. [35]

    Soft actor-critic for discrete action settings,

    P. Christodoulou, “Soft actor-critic for discrete action settings,” 2019. [Online]. Available: https://arxiv.org/abs/1910.07207

    2019 arXiv

  28. [2009]

    Available: https://api.semanticscholar.org/CorpusID: 125587700

    [Online]. Available: https://api.semanticscholar.org/CorpusID: 125587700

  29. [2016]

    Available: https://arxiv.org/abs/1604.07316

    [Online]. Available: https://arxiv.org/abs/1604.07316

    arXiv

  30. [2019]

    Available: https://arxiv.org/abs/1910.05449

    [Online]. Available: https://arxiv.org/abs/1910.05449

    1910 arXiv

  31. [2023]

    Available: https://arxiv.org/abs/2306.15968

    [Online]. Available: https://arxiv.org/abs/2306.15968

    arXiv

Pith tools

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