Knowledge Distillation and Reinforcement Learning in a Human–Machine Collaboration Delivery System With a Robotic Arm

Clinical Snapshot

Key Findings

• P Value: Not reported

• Effect Size: Decision RL Model accuracy: 99.17%; Merged Model accuracy: 99.92%

• Primary Outcome: Task delivery accuracy of reinforcement learning models in a human-robot collaboration object delivery scenario

• Nnt Or Sensitivity: Not applicable in traditional clinical sense; no sensitivity/specificity, NNT, or hazard ratio reported. No comparative accuracy figure for the baseline conventional path planning method is provided, precluding calculation of relative improvement.

• Confidence Interval: Not reported

Clinical Application

Technically feasible as a laboratory proof-of-concept. Real-world clinical deployment would require: extensive safety validation including failure mode and effects analysis (FMEA); real-world trials across diverse environments and user populations; regulatory approval pathways; integration with existing hospital infrastructure; and staff training. Computational requirements for real-time RL inference in clinical settings are not characterised. In Australia, robotic medical devices and assistive technologies are regulated by the Therapeutic Goods Administration (TGA) under the medical devices framework if used in clinical care. Robotic systems used in aged care would also fall under the Aged Care Quality and Safety Commission standards. The RACGP and relevant specialist colleges have not issued guidance on clinical robotic delivery systems. PBS listing is not relevant to this technology at this stage. Workplace health and safety obligations under Safe Work Australia would apply to any hospital deployment. Australian aged care and disability sectors (NDIS) represent plausible future application domains, but no Australian-specific validation data exists. This study is not directly applicable to a clinical patient population. Potential future applications include hospital logistics and supply delivery robots, aged care and disability support robotic assistants, rehabilitation robotics requiring precise object handover, and surgical or procedural instrument delivery systems. All such applications would require substantial additional validation.

Abstract

Robotic arms are widely used in various aspects of human-robot collaboration. The primary goal of this study is to explore the usability of robotic arms for delivering objects to humans in dynamic environments. Traditional robotic arms often face limitations in path planning, such as difficulties adapting to dynamic environments and complex developmental processes. To overcome these challenges, this study employs reinforcement learning (RL) to train four models-the Approach RL Model, Delivery RL Model, Decision RL Model, and Merged Model-as alternatives to conventional path planning control. Typically, there exists a significant discrepancy between simulated data and real-world features. Although image segmentation can substantially reduce the gap between virtual and real environments, notable differences remain in hand features. Therefore, to further bridge the simulation-to-reality gap, this study applies CycleGAN to transform real hand features into virtual hand features, thereby enhancing the model's transferability. Experimental results show that the Decision RL Model achieved an accuracy of 99.17%, while the Merged Model achieved 99.92%. The proposed method effectively improves the stability and accuracy of human-robot collaboration in complex scenarios. Overall, this study validates the feasibility of integrating RL, image segmentation, and image translation techniques, offering a scalable and efficient task-solving solution for robotic arms in highly dynamic application domains.

References

1. 1.Kuo, P.-H., Feng, P.-H., Chang, C.-W., Lin, Y.-S., Chiu, Y.-C., & Chen, B.-Y. (2026). Knowledge distillation and reinforcement learning in a human–machine collaboration delivery system with a robotic arm. IEEE Transactions on Cybernetics. https://doi.org/10.1109/TCYB.2026.3668072