REVIEW 4 major objections 4 minor 17 references
Conditional Prediction by Simulation for Automated Driving
T0 review · 4 major / 4 minor · reviewed 2026-08-09 · deepseek-v4-flash
Pith's one-line read When a self-driving car changes its plan, the simulation changes how other drivers are predicted to react.
desk verdict A plausible integration of AIRL into conditional prediction, but the central claim rests on a single qualitative example. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The mechanism is a closed-loop traffic simulation in which a graph-based behavior model, trained with Adversarial Inverse Reinforcement Learning, controls every non-ego vehicle. Agent-centric observations encode vehicles and road elements within 30 meters as polylines, passed through message passing and cross-attention to produce the next acceleration and steering angle; conditioning is implemented by overriding the ego vehicle's executed action with its candidate plan at each simulation step. The AIRL reward is modified with a constant survival bonus and action noise during discriminator training to keep the simulated drivers on track, and Proximal Policy Optimization with Generalized Advantage Estimation updates the shared policy.
What would settle it
Record or collect real-world cases in which a driver takes an alternative action, such as braking at an all-way stop instead of proceeding first, run the conditional simulation with that action as the conditioned plan, and compare the simulated trajectories of the other agents against the ground-truth trajectories from those cases. A quantitative divergence on such a conditional benchmark would show that the model generalizes to conditioned plans only in appearance, not in distribution.
Extended reading notes
Core claim
The central claim is that a single behavior policy trained via Adversarial Inverse Reinforcement Learning and executed simultaneously for all vehicles in a stepwise traffic simulation is sufficient to generate conditional predictions that capture bidirectional interactions between an automated vehicle and its surroundings. To condition a forecast on a candidate plan, the automated vehicle's state is updated along the plan during the rollout while all other agents follow the learned policy; because each predicted state depends on the previous states of every vehicle, the other drivers' simulated trajectories shift in response to the plan, and the plan itself can in principle adapt to those reactions. The paper demonstrates this with an all-way-stop intersection where replacing the ego vehicle's motion with constant braking changes the predicted progress of the oncoming vehicle and the vehicle behind it, while leaving uninvolved vehicles unaffected.
Load-bearing premise
The learned driver policy must respond realistically to ego trajectories that differ from what the training drivers actually did, even when those conditioned plans push the situation outside the distribution of the recorded demonstrations.
Editorial extensions
If this is right
- A planner can score candidate plans by simulating each one and using the resulting conditional predictions to choose maneuvers that other drivers are predicted to accommodate.
- Because all agents share the same learned policy and react to each other's updated states, the predictions remain scene-consistent over a 10-second horizon, avoiding the frozen-robot failure of purely reactive prediction.
- The stepwise rollout lets the ego plan be revised during prediction, so planning strategies can be built as trees or receding-horizon searches rather than single fixed trajectories.
- Conditioning only changes the agents that are actually affected by the ego vehicle's alternative plan, so the model does not need to be retrained for each new query.
- Training in closed-loop simulation with AIRL yields markedly lower collision and off-track rates than behavior cloning after 10 seconds, which is what makes the conditional rollouts usable for planning.
Reading between the lines
- The paper validates conditioning qualitatively on one intersection; a natural next step is a quantitative benchmark that compares conditional predictions against ground-truth cases where drivers actually took the conditioned action.
- Because the interaction graph and shared-policy design do not depend on which agent is conditioned, the same machinery could simulate hypothetical actions of multiple agents or of non-vehicle road users.
- If the conditional forecasts generalize beyond the training distribution, a planner could use them to decide when to assert right-of-way, proceeding when the simulation predicts that other drivers will yield.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes a conditional motion prediction framework for automated driving. A behavior policy is trained via Adversarial Inverse Reinforcement Learning on the INTERACTION dataset and then used as a closed-loop simulator in which all vehicles execute the same shared policy. To make conditional predictions, the automated vehicle's actions are replaced by a manually defined plan, and the other agents' reactions are simulated. The authors report unconditional prediction performance (RMSE, collision rate, off-track rate) showing AIRL outperforms behavior cloning, and provide a single qualitative example of conditional prediction at an all-way stop intersection.
Significance. The contribution is potentially significant: if the learned policy produces realistic responses to alternative AV plans, it would enable evaluating cooperative strategies and mitigate the freezing robot problem. The unconditional results are encouraging, particularly the low collision and off-track rates, and the use of AIRL for a closed-loop model is a sensible direction. However, the paper's central conditional claim is not quantitatively evaluated: the single example is illustrative only, with no conditional metrics, baselines, or error bars, and the paper overclaims dynamic plan adaptation. These gaps must be addressed before the contribution can be regarded as demonstrated.
major comments (4)
- [Section 3.1 (Conditional Prediction paragraph, Figure 4) and Section 2.4] The paper's central claim—that the framework produces meaningful conditional predictions for alternative AV plans—is supported only by a single qualitative example. There are no conditional metrics, no baselines (e.g., a reactive model that ignores the AV plan), no error bars, and no systematic evaluation across scenarios. Because the AIRL policy is trained in closed-loop settings where all agents follow the same policy, imposing an external plan as in Figure 4(b) places other agents out of distribution; the paper does not test whether the resulting reactions are realistic or consistent with human behavior. Without such evidence, the conditional-prediction claim is not demonstrated. Concretely, the authors should add quantitative conditional evaluations: in several scenarios, compare conditional rollouts against a reactive baseline, measure the distribution of predicted actions under conditioning against expert action statistics, and report how often the conditioned plan leads to collisions or off-track events rather than removing those vehicles.
- [Abstract and Section 2.4] The abstract states that "our approach allows the candidate trajectories to adapt dynamically during the prediction rollout," but Section 2.4 explicitly says "we leave this for future research and use manually defined fixed paths for the A V during prediction." The Introduction similarly claims the framework allows the A V to adapt its planned trajectory dynamically. This dynamic adaptation is neither implemented nor tested in the manuscript. The authors should either implement a reactive planner or revise the abstract and introduction to match the fixed-path implementation actually used.
- [Section 3 (first paragraph of Experiments)] The sentence "Vehicles that reach the end of their assigned route, leave the track, or collide are removed from the scene and the simulation is continued with the remaining vehicles" introduces a bias into the reported statistics. Removing vehicles that leave the track or collide excludes the most severe failures from the RMSE computation, potentially understating errors. This is especially problematic for conditional prediction, where a conditioned plan that induces collisions would be silently censored rather than reported as a failure. Please report results with and without removal, or at least analyze how this design choice affects the RMSE and the collision/off-track rates, and specify how RMSE is computed for vehicles that are removed before the prediction horizon ends.
- [Section 2.2, Eq. (4) and the Modifications paragraph] The claim that the reward offset c "promot[es] survival without changing the optimal behavior with respect to the discriminator model" requires qualification in the presence of terminal states. If episodes terminate on off-track or collision, adding a positive constant c to every step changes the value of surviving and can alter the optimal policy by making the agent more averse to risky states. The manuscript does not provide a theoretical justification or empirical evidence that the learned policy's behavior is unaffected by c. Please clarify this point or justify the choice of c more rigorously.
minor comments (4)
- [Section 3.1 (Table 2)] The table reports RMSE after 10 s, but the text does not define how RMSE is computed (e.g., mean Euclidean position error over time? per vehicle? how are vehicles matched between prediction and ground truth?). Please specify the metric precisely.
- [Section 3.1 (Conditional Prediction paragraph)] The paper mentions "Several example scenarios are available at [website]" but the main text includes only one conditional example. Including additional conditional examples or quantitative summaries in the paper would strengthen the presentation of the central contribution.
- [Figure 4] The ground-truth panel (c) corresponds to the unconditional prediction (a), not to the conditional prediction (b). The authors should clarify that the ground truth does not validate the counterfactual, and that the conditional prediction is an illustration rather than a quantitatively evaluated result.
- [Section 2.3] The policy decoder outputs the mean and standard deviation of the next acceleration and steering angle, but the distribution family is not stated. Please specify (e.g., independent Gaussians) so that Eq. (3) and the action likelihood are unambiguous.
Circularity Check
No significant circularity: the conditional simulation directly executes a policy trained on demonstrations, with no fitted parameter renamed as prediction.
full rationale
The paper's claimed derivation is self-contained: AIRL trains a shared behavior policy on naturalistic demonstrations, and at inference the surrounding agents execute this policy in a closed-loop simulation while the AV follows a manually fixed plan; the resulting scene evolution is the conditional prediction. Each link is verifiable from the paper's own description. The policy is not fitted to the conditional example in Figure 4, the action replacement for vehicle ID 6 is an input intervention rather than a fitted target, and the reported RMSE, collision, and off-track metrics are unconditional evaluations against held-out INTERACTION data. The same-authors citations [7] and [10] supply architectural and training modifications, but the present experiments and the conditional example are generated by the current model, so those citations are background support rather than a load-bearing reduction. The hand-tuned reward offset c=5 affects survival tendency but does not determine the conditional responses shown. The conditional-prediction claim is demonstrated only qualitatively (Figure 4), and out-of-distribution plan generalization is not quantitatively validated; this is a correctness and validation weakness, not circularity, because no target output is encoded in the inputs or fitted parameters. The paper also explicitly leaves dynamic plan adaptation to future work while the abstract mentions it as a capability, but this is a consistency gap rather than a circular step.
Assumptions & free parameters
free parameters (1)
- surrogate reward constant c =
5
assumptions (4)
- standard math PPO/GAE and AIRL converge to a policy that maximizes the reconstructed reward and matches expert demonstrations.
- domain assumption Expert demonstrations in INTERACTION capture realistic human driving behavior, and the recovered reward generalizes to novel scenes.
- domain assumption Every vehicle's future route is known or can be inferred online from the dataset.
- ad hoc to paper Removing vehicles that leave the track or collide, then continuing simulation, keeps the evaluation valid.
Cite this review
Pith. "Pith review of Conditional Prediction by Simulation for Automated Driving." pith.science (2026). https://pith.science/paper/35IN6KRQ
@misc{pith2026250203286,
author = {Pith},
title = {Pith review of: Conditional Prediction by Simulation for Automated Driving},
year = {2026},
howpublished = {\url{https://pith.science/paper/35IN6KRQ}},
note = {Machine review of arXiv:2502.03286}
}
read the original abstract
Modular automated driving systems commonly handle prediction and planning as sequential, separate tasks, thereby prohibiting cooperative maneuvers. To enable cooperative planning, this work introduces a prediction model that models the conditional dependencies between trajectories. For this, predictions are generated by a microscopic traffic simulation, with the individual traffic participants being controlled by a realistic behavior model trained via Adversarial Inverse Reinforcement Learning. By assuming various candidate trajectories for the automated vehicle, we generate predictions conditioned on each of them. Furthermore, our approach allows the candidate trajectories to adapt dynamically during the prediction rollout. Several example scenarios are available at https://conditionalpredictionbysimulation.github.io/.
Figures
Figures from the paper (1 more)
Reference graph
Works this paper leans on
-
[1]
Motionlm: Multi-agent motion forecasting as language modeling,
A. Seff, B. Cera, D. Chen, M. Ng, A. Zhou, N. Nayakanti, K. S. Refaat, R. Al-Rfou, and B. Sapp, “Motionlm: Multi-agent motion forecasting as language modeling,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 8579–8590
work page 2023
-
[2]
Mtr++: Multi-agent motion prediction with symmetric scene modeling and guided intention querying,
S. Shi, L. Jiang, D. Dai, and B. Schiele, “Mtr++: Multi-agent motion prediction with symmetric scene modeling and guided intention querying,” IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024
2024
-
[3]
Scenemotion: From agent-centric embeddings to scene-wide forecasts,
R. Wagner, ¨O. S. Tas, M. Steiner, F. Konstantinidis, H. K¨ onigshof, M. Klemp, C. Fernandez, and C. Stiller, “Scenemotion: From agent-centric embeddings to scene-wide forecasts,” in International Conference on Intelligent Transportation Systems (ITSC), 2024
work page 2024
-
[4]
Unfreezing the robot: Navigation in dense, interacting crowds,
P. Trautman and A. Krause, “Unfreezing the robot: Navigation in dense, interacting crowds,” in 2010 IEEE/RSJ International Conference on Intelligent Robots and Systems. IEEE, 2010, pp. 797–803
work page 2010
-
[5]
Identi- fying driver interactions via conditional behavior prediction,
E. Tolstaya, R. Mahjourian, C. Downey, B. Vadarajan, B. Sapp, and D. Anguelov, “Identi- fying driver interactions via conditional behavior prediction,” in 2021 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2021, pp. 3473–3479
work page 2021
-
[6]
Tree-structured policy planning with learned behavior models,
Y. Chen, P. Karkus, B. Ivanovic, X. Weng, and M. Pavone, “Tree-structured policy planning with learned behavior models,” in 2023 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2023, pp. 7902–7908
work page 2023
-
[7]
F. Konstantinidis, M. Sackmann, U. Hofmann, and C. Stiller, “Modeling interaction-aware driving behavior using graph-based representations and multi-agent reinforcement learning,” in 2023 IEEE International Intelligent Transportation Systems Conference (ITSC). IEEE, 2023
work page 2023
-
[8]
Modeling driver behavior using adversarial inverse reinforcement learning,
M. Sackmann, H. Bey, U. Hofmann, and J. Thielecke, “Modeling driver behavior using adversarial inverse reinforcement learning,” in 2022 IEEE Intelligent Vehicles Symposium (IV). IEEE, 2022, pp. 1683–1690
work page 2022
Show all 17 references
-
[9]
Feedback in imitation learning: The three regimes of covariate shift,
J. Spencer, S. Choudhury, A. Venkatraman, B. Ziebart, and J. A. Bagnell, “Feedback in imitation learning: The three regimes of covariate shift,” arXiv preprint arXiv:2102.02872, 2021
2021 arXiv
-
[10]
Graph-based adversarial imitation learning for predicting human driving behavior,
F. Konstantinidis, M. Sackmann, U. Hofmann, and C. Stiller, “Graph-based adversarial imitation learning for predicting human driving behavior,” in 2024 IEEE Intelligent Vehicles Symposium (IV). IEEE, 2024, pp. 857–864
2024
-
[11]
Learning robust rewards with adversarial inverse reinforcement learning,
J. Fu, K. Luo, and S. Levine, “Learning robust rewards with adversarial inverse reinforcement learning,” arXiv preprint arXiv:1710.11248, 2017
2017 arXiv
-
[12]
High-dimensional continuous control using generalized advantage estimation,
J. Schulman, P. Moritz, S. Levine, M. Jordan, and P. Abbeel, “High-dimensional continuous control using generalized advantage estimation,” arXiv preprint arXiv:1506.02438, 2015
2015 arXiv
-
[13]
Proximal policy opti- mization algorithms,
J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Proximal policy opti- mization algorithms,” arXiv preprint arXiv:1707.06347, 2017
2017 arXiv
-
[14]
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,”Advances in neural information processing systems, vol. 30, 2017
2017
-
[15]
Multiple model unscented kalman filtering in dynamic bayesian networks for intention estimation and trajectory prediction,
J. Schulz, C. Hubmann, J. L¨ ochner, and D. Burschka, “Multiple model unscented kalman filtering in dynamic bayesian networks for intention estimation and trajectory prediction,” in 2018 21st International Conference on Intelligent Transportation Systems (ITSC). IEEE, 2018, pp...
2018
-
[16]
Adam: A method for stochastic optimization,
D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980, 2014
2014 arXiv
-
[17]
Interaction dataset: An international, adversarial and cooperative motion dataset in interactive driving scenarios with semantic maps,
W. Zhan, L. Sun, D. Wang, H. Shi, A. Clausse, M. Naumann, J. Kummerle, H. Konigshof, C. Stiller, A. de La Fortelle et al., “Interaction dataset: An international, adversarial and cooperative motion dataset in interactive driving scenarios with semantic maps,” arXiv preprint ar...
1910 arXiv
Reviewed August 9, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.