Does Knowledge About Perceptual Uncertainty Help an Agent in Automated Driving?

arXiv:2502.11864 · cs.CV, cs.RO · Submitted 2025-02-17 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom, and with me are Jane, Lu, senior AI researcher at Tsinghua, Meng, lead engineer at a mysterious AI startup and Lalam, the in-house Large Language Model.

Jane: Today's paper: "Does Knowledge About Perceptual Uncertainty Help an Agent in Automated Driving?".

Tom: The gist: Agents in real-world scenarios like automated driving deal with uncertainty in their environment, in particular due to perceptual uncertainty.

Jane: First, who's behind it and why it matters.

Paper summary: Tom: So we’ve looked at how grabbing uncertainty information helps an agent adapt its driving behavior, which is what the authors of "Does Knowledge About Perceptual Uncertainty Help an Agent in Automated Driving?" are really asking about. It's Natalie Grabowsky and Annika Mutze doing this work.

Jane: The paper suggests that while having perfect perception is ideal, knowing when your perception might be wrong can actually help the agent perform its goal—driving fast without colliding—more effectively than just relying on potentially misleading data alone.

Lu: It implies that in real-world AI, the value isn't just in having the most accurate input, but in having a mechanism to signal when that input is questionable so the decision-making process can adjust accordingly.

Meng: For practical engineering, this means we need better ways to estimate perception uncertainty because it directly impacts whether an autonomous vehicle drives defensively or aggressively under specific conditions.

Lalam: It gives us a path toward building systems where the AI doesn't just blindly follow what its sensors report, but intelligently weighs the confidence in those reports against the immediate driving task.

Tom: It’s about moving from simple input-output learning to a system that understands its own limitations, which is pretty significant for scaling up autonomous systems.

Conclusion: Segment: Conclusion**

Tom: So, to wrap up on this study by Grabowsky and Mutze, they’re asking if knowing about perception uncertainty helps an agent drive better or worse. They used a driving simulator called CARLA and the PPO algorithm to test how the agent acts when it has or doesn't have info about its sensors being unreliable.

Jane: It boils down to this: if you only trust what your sensors tell you, the AI might get stuck in overly defensive modes because it’s not sure what’s actually out there on the road. But when they give the agent a signal—an indicator—that its perception is shaky, it learns to adjust its speed and distance more intelligently.

Lu: What this means for AI design is that just having high-fidelity data isn't enough; you need that uncertainty metric built right into how the AI makes decisions, like feeding it directly into the observation space.

Meng: From an engineering standpoint, this suggests we can build systems that are proactive instead of just reactive when sensor noise hits. It’s about making sure the AI knows *when* it should slow down or change its plan because something looks weird.

Lalam: That’s huge for how we think about building these systems; it moves the AI away from just being a fast reaction machine and toward a more calibrated decision-maker that respects its own limitations.

Tom: It shifts the focus from just making the perception perfect to making the decision-making robust even when perception isn't perfect. This whole paper is really about how we teach autonomous agents to handle real-world messiness without breaking their driving goals or crashing. Next up, we’re going to look at what exactly they did in their experiments with those numbers and see how they actually behaved during the test runs.

Institute of Mathematics, Technical University of Berlin · Department of Mathematics, University of Wuppertal · Faculty of Computer Science, Ruhr University Bochum

cs.CV, cs.RO

Submitted: 2025-02-17

Updated: 2026-10-09

Comments: 11 pages, 9 figures, Appendix

Code: https://github.com/nagrab/Does-Knowledge-About-Perceptual-Uncertainty-Help-anAgent-in-Automated-Driving

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 74/100

The gist: The gist: Agents in real-world scenarios like automated driving deal with uncertainty in their environment, in particular due to perceptual uncertainty.

Key concepts

Perceptual Uncertainty: This refers to the doubt or lack of certainty in what an agent perceives from its sensors, like a camera. It models situations where the sensor data might be flawed or ambiguous, which is common in real-world driving scenarios.
Proxy Task: This is a simplified goal used for testing. In this research, the agent was rewarded for driving as fast as possible without crashing into other road users. This task allowed researchers to observe how uncertainty affects decision-making without focusing solely on the final, complex autonomous driving goal.
Observation Space Augmentation: This involves modifying what the agent 'sees' or receives. The researchers specifically added a 'Uncertainty component' to the observation space. This component provides information about which specific perceptual perturbation is currently being experienced, giving the agent extra context beyond just raw visual data.

Terminology

Calibration — 1× in this paper · explained in 7 episodes
Reinforcement Learning — 1× in this paper · explained in 4 episodes

Summary

The gist: Agents in real-world scenarios like automated driving deal with uncertainty in their environment, in particular due to perceptual uncertainty.

Motivation and Problem

Agents in real-world scenarios like automated driving deal with uncertainty in their environment, particularly due to perceptual uncertainty<ref:2502.11864#pg2>. Although reinforcement learning is dedicated to autonomous decision-making under uncertainty, these algorithms are typically not informed about the uncertainty currently contained in their environment<ref:2502.11864#pg3>. Uncertainty estimation for perception itself is typically directly evaluated in the perception domain, e.g., in terms of false positive detection rates or calibration errors based on camera images<ref:2502.11864#pg2>. Its use for deciding on goal-oriented actions remains largely unstudied<ref:2502.11864#pg2>. Agents that base their decision-making on this perception may end up in life-threatening failures<ref:2502.11864#pg3>.

Experimental Setup

The study investigates how an agent’s behavior is influenced by an uncertain perception and how this behavior changes if information about this uncertainty is available<ref:2502.11864#pg2>. The proxy task considered is where the agent is rewarded for driving a route as fast as possible without colliding with other road users<ref:2502.11864#pg3>. Controlled experiments introduce uncertainty in the observation space by perturbing the perception of the given agent while informing the latter<ref:2502.11864#pg2>. The observation space is defined by three parts: Vision component, Non-visual component, and Uncertainty component<ref:2502.11864#pg2>. The observation function maps state S directly into the observation space O, where O = Cvision × Cnon-vis × Cuncertainty<ref:2502.11864#pg3>.

Key Findings from Experiments

The experiments reveal distinct behavioral patterns based on the availability of uncertainty information<ref:2502.11864#pg2>.

Unreliable perception typically results in defensive driving behavior<ref:2502.11864#pg3>. The agent trained in scenario 2 (perturbed perception without uncertainty information) acts generally defensively to avoid collisions, and performs ‘brake’ more frequently than in the unperturbed case<ref:2502.11864#pg2>.
When perceptual uncertainty is incorporated into the observation space (scenario 3), the agent adapts to the specific situation and in general accomplishes its task faster while, at the same time, accounting for risks<ref:2502.11864#pg3>.
Under a perturbed perception it acts more defensively while increasing the pace when the uncertainty information states that the perception is correct<ref:2502.11864#pg3>. For instance, between 1500 and 1800 time steps, the agent knows only the perception w.r.t. the following vehicle is perturbed (XEVV), but not w.r.t. the front vehicle, in which case it reduces its distance to the front vehicle to about 10 and 15 meters<ref:2502.11864#pg3>.

Conclusion

The paper concludes that an agent trained on perturbed perception with uncertainty information adapts to the given scenario<ref:2502.11864#pg3>. The agent is able to finish its task in all three scenarios when tested on the training scenario, but fails consistently when evaluated on a perturbed perception but trained with a correct one<ref:2502.11864#pg2>.

Related Work Context

The research addresses an open question in RL on how ambiguous and noisy data can be handled by RL approaches<ref:2502.11864#pg3>, as well as the open question in CV how relevant uncertainty estimation for perception is in AD<ref:2502.11864#pg3>. It examines the interface between RL and perceptual uncertainty estimation<ref:2502.11864#pg2> and explores how this information influences subsequent actions of an RL agent<ref:2502.11864#pg2>. The study contrasts its findings with prior work that assumes accurate observations, investigating how the agent’s behavior is influenced when it is informed about the presence or absence of uncertainty in the perception<ref:2502.11864#pg3>.

Methodology Details

The learning task uses the on-policy PPO algorithm<ref:2502.11864#pg2> interacting with a driving simulator environment like CARLA<ref:2502.11864#pg2>. The agent’s objective is to find a policy that maximizes the expected discounted cumulative reward, defined by a time decay component rewarding momentary speed in the direction of the end point<ref:2502.11864#pg3>. The observation space dimension increases when uncertainty is added, specifically through the Cuncertainty component which stores one-hot encoded information about which of the four perceptual perturbations is currently being experienced<ref:2502.11864#pg3>. The action space A = [−1, 1] represents accelerating or braking intensity, and a modified action at t is given by an inertia model to simulate realistic driving dynamics<ref:2502.11864#pg4>.

Summary of Contributions

The paper summarizes its contributions as follows:

We provide an experimental setup for RL under an uncertain segmentation-based perception. This uncertain perception is complemented with indicators of uncertainty, mimicking optimal uncertainty estimation.

We investigate the role of perceptual uncertainty estimation in enhancing the downstream RL task of AD.

"Our experiments reveal that unreliable perception typically results in defensive driving behavior. However, when perceptual uncertainty is incorporated into the observation space, the agent learns to adapt more effectively to changing situations."

References Used

[1] M. Bojarski and et al., “End to end learning for self-driving cars,” arXiv:1604.07316, 2016.

[2] Z. Chen and X. Huang, “End-to-end learning for lane keeping of self-driving cars,” in 2017 IEEE Intelligent Vehicles Symposium (IV), pp. 1856–1860, 2017.

[3] T. Agarwal, H. Arora, and J. Schneider, “Learning urban driving policies using deep reinforcement learning,” in 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pp. 607–614, 2021.

[4] J. Chen, B. Yuan, and M. Tomizuka, “Model-free deep reinforcement learning for urban autonomous driving,” in 2019 IEEE Intelligent Transportation Systems Conference (ITSC), pp. 2765–2771, 2019.

[5] E. Perot, M. Jaritz, M. Toromanoff, and R. de Charette, “End-to-end driving in a realistic racing game with deep reinforcement learning,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, July 2017.

[6] M. Jaritz, R. de Charette, M. Toromanoff, E. Perot, and F. Nashashibi, “End-to-end race driving with deep reinforcement learning,” in 2018 IEEE International Conference on Robotics and Automation (ICRA), pp. 2070–2075, 2018.

[7] R. Sutton and A. Barto, “Reinforcement learning: An introduction,” IEEE Transactions on Neural Networks, vol. 9, no. 5, pp. 1054–1054, 1998.

[8] M. Toromanoff, E. Wirbel, and F. Moutarde, “End-to-end model-free reinforcement learning for urban driving using implicit affordances,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 7153–7162, 2020.

[9] F. Ye, X. Cheng, P. Wang, C.-Y. Chan, and J. Zhang, “Automated lane change strategy using proximal policy optimization-based deep reinforcement learning,” in 2020 IEEE Intelligent Vehicles Symposium (IV), pp. 1746–1752, 2020.

[10] H. Park, “Trajectory planning for autonomous vehicle using iterative reward prediction in reinforcement learning,” arXiv:2404.12079, 2024.

[11] A. Sootla, A. I. Cowen-Rivers, T. Jafferjee, Z. Wang, D. H. Mguni, J. Wang, and H. Ammar, “Saute RL: Almost surely safe reinforcement learning using state augmentation,” in Proceedings of the 39th International Conference on Machine Learning, vol. 162 of Proceedings of Machine Learning Research, pp. 20423–20443, PMLR, July 2022.

[12] G. Kahn, A. Villaflor, V. Pong, P. Abbeel, and S. Levine, “Uncertainty-aware reinforcement learning for collision avoidance,” CoRR, vol. abs/1702.01182, 2017.

[13] B. Lutjens, M. Everett, and J. P. How, “Safe reinforcement learning ¨ with model uncertainty estimates,” in 2019 International Conference on Robotics and Automation (ICRA), pp. 8662–8668, 2019.

[14] H. Zhang and Y.

Improvements for AI systems

textbf Real-time uncertainty-aware policy adaptation: The improved RL agent can adapt to a specific situation because it receives information about this uncertainty directly to the observation space, leading to better performance than agents that only receive perturbed observations.
textbf Enhanced safety through informed risk assessment: The agent can achieve safer driving by using the uncertainty information, as the paper shows that when perception is unreliable, an agent typically results in defensive driving behavior, and providing uncertainty information allows it to adapt to the specific situation and in general accomplishes its task faster while, at the same time, accounting for risks.
textbf Dynamic risk management during perception shifts: The system can handle spontaneous changes in perception quality by exhibiting adaptation; specifically, an agent trained on perturbed perception with uncertainty information shows that it can adapt to the given scenario, even exhibiting inertia in terms of reaction time needed to adapt to the current perturbation case.
textbf Optimized driving strategy based on perceived certainty: The agent's action selection is improved by integrating the uncertainty component, as the observation space is defined as O = Cvision × Cnon-vis × Cuncertainty, allowing it to modulate its speed and braking decisions based on how certain it is about the visual input.