AI xPress

AI and computer science research.

Back to Foragentis

Robotics & Embodied AI

Human guidance during training improved the efficiency and performance of deep reinforcement learning for autonomous driving, a 2022 study suggests

Robotics & Embodied AI November 2, 2026

A robotic prosthetic hand on a workbench

In a 2022 paper in the journal Engineering, researchers at Nanyang Technological University tested a way to put a person inside the training process of a deep reinforcement learning system. Deep reinforcement learning is a machine learning method in which an agent learns by trial and error from rewards. The system learned to drive in an end-to-end case, meaning it went directly from inputs to driving actions. Human participants could take control in real time when the agent chose an unreasonable action. In experiments with 40 subjects, the results suggest that this human guidance improved the training efficiency and performance of the algorithm compared with other learning approaches. The improvement did not require participants to have any particular expertise or experience.

What did the researchers do?

They developed a real-time human-guidance method called Hug-DRL. A control transfer mechanism let a person intervene and correct the agent during training. The researchers also modified an actor-critic network, the part of the system that learns a policy for what action to take and a value for how good a state is, so it could accept the human corrections.

What did the study find?

The results suggest that real-time human guidance can improve the efficiency and performance of DRL training in the autonomous driving case. The gains did not depend on selecting participants for special skills or experience.

Who was studied?

Forty subjects took part in human-in-the-loop experiments. The abstract describes an end-to-end autonomous driving case. It does not say whether the driving happened in simulation or on real roads.

Limits and what's next

The study was a human-in-the-loop experiment with 40 subjects in an end-to-end autonomous driving case. The abstract does not name the driving data or environment used. Because the abstract reports one study and no independent replication, the finding is tied to this sample and setting.

Source: Jingda Wu et al., “Toward Human-in-the-Loop AI: Enhancing Deep Reinforcement Learning via Real-Time Human Guidance for Autonomous Driving,” Engineering, 2022.