Robust Reinforcement Learning for Autonomous Systems

Methodology

• Adversarial Analysis – systematic evaluation of how attacks influence deep reinforcement learning models.
• Defense Mechanisms – design and testing of ensemble defense approaches that combine multiple countermeasures.
• Robustness Evaluation – benchmarking model resilience in simulated and controlled real-world-like environments to ensure safety and reliability.
   

Selected Research Results

Our research on robust reinforcement learning focuses on understanding how intelligent agents behave under uncertainty and adversarial influence, and how these vulnerabilities can be addressed without compromising policy efficiency or deployment feasibility. The following lines of work highlight key scientific insights emerging from our recent studies.

Improving Robustness with Ensemble Defense Mechanisms:
A collaboration between Adithya Mohan, Dominik Rößle, Daniel Cremers and Torsten Schön investigates how multiple lightweight defense mechanisms can be combined to protect deep reinforcement learning agents from adversarial perturbations in autonomous-driving scenarios. The proposed approach aggregates three complementary preprocessing methods—randomized noise injection, autoencoder-based reconstruction, and PCA filtering - into a single inference-time ensemble. As shown in controlled Highway-env and Merge simulations, the ensemble substantially improves reward recovery and reduces collision rates compared to individual defenses, all without retraining or modifying the underlying policy network. This demonstrates that modular, inference-only defenses can significantly enhance the safety of deployed agents operating in the presence of sensor noise or malicious input manipulation.

Quantifying Robustness Across Driving Scenarios:
Our experiments across different simulated traffic tasks show that adversarial influence affects driving behavior in distinct ways depending on scene structure.
Highway scenarios, with higher speeds and denser interactions, exhibit sharper degradation under attack and larger performance variance, whereas merge scenarios reveal how simpler decision spaces still expose agents to critical misjudgments when perturbed. By benchmarking defenses across settings with varying complexity, we gain a clearer picture of how robustness mechanisms scale and where they may require adaptation before real-world integration.

Toward Trustworthy RL for Intelligent Mobility:
Together, these results form a methodological foundation for designing reinforcement learning systems that remain functional under uncertainty and malicious influences. The findings contribute to a broader goal: enabling trustworthy, interpretable, and safe AI agents for future autonomous mobility applications.
 

Funding

The research is funded by the Hightech Agenda Bayern.

Impact

This research advances the development of resilient and transparent AI models for autonomous driving. By focusing on robustness in reinforcement learning, CVIMS contributes to safer, more predictable behavior of intelligent agents, which is a key requirement for public trust and large-scale deployment of autonomous mobility systems in Ingolstadt and beyond.

Publications

• Advancing Robustness in Deep Reinforcement Learning with an Ensemble Defense Approach
• Toward Robust Agents: A Survey of Adversarial Attacks and Defenses in Deep Reinforcement Learning