Reinforcement Learning Applications in Autonomous Mobility Systems

Reinforcement learning (RL) has emerged as a powerful paradigm for sequential decision-making in autonomous mobility systems, enabling vehicles and fleets to learn optimal behaviours through trial and reward signals rather than explicit programming. In individual vehicle control, RL algorithms have been applied to tasks such as adaptive cruise control, collision avoidance and energy-efficient routing, where continuous action spaces and real-time feedback are critical. At the fleet level, multi-agent RL supports coordinated dispatching, vehicle rebalancing and dynamic pricing in mobility-on-demand services, promoting system-wide efficiency and passenger service quality. Advances in deep neural policy representations have underpinned breakthroughs in high-dimensional state handling, permitting autonomous vehicles to process sensor streams and traffic data to inform policy updates. Recent work also explores hierarchical RL for decomposing complex mobility objectives—such as ride-pool matching and zone-level supply balancing—into tractable sub-tasks. Together, these developments are shaping urban transport ecosystems that self-optimise in response to fluctuating demand patterns, traffic conditions and energy constraints, with broad implications for sustainability and equity in shared mobility.

Research from Nature Portfolio

No recent Nature Portfolio content available.

Reinforcement Learning Applications in Autonomous Mobility Systems publication trend

The graph below shows the total number of articles in reinforcement learning applications in autonomous mobility systems across all publications each year (not limited to Nature Index journals).

Technical terms

Reinforcement learning: A machine-learning paradigm in which an agent learns to make decisions by taking actions in an environment to maximise cumulative reward.

Markov decision process: A formal framework describing environments in terms of states, actions, transition probabilities and reward functions, underlying most RL formulations.

Deep Q-network: A neural-network approximation of the Q-value function in which the expected reward of state–action pairs is iteratively estimated.

Reward function: A mapping from state and action to a scalar value that guides the agent towards desirable outcomes during learning.

Vehicle rebalancing: The optimisation of relocating idle vehicles to regions with anticipated high demand to reduce passenger wait times and fleet imbalance.

Exploration–exploitation trade-off: The dilemma in RL between exploring new actions to discover high reward and exploiting known actions to maximise immediate gain.

References

1. Dynamic Fleet Management With Rewriting Deep Reinforcement Learning. IEEE Access (2020).
2. An Application of Reinforced Learning-Based Dynamic Pricing for Improvement of Ridesharing Platform Service in Seoul. Electronics (2020).
3. A survey on applications of reinforcement learning in spatial resource allocation. Computational Urban Science (2024).