Human Robot Interaction

This work focused on the time delay issues in human robot interactions, especially in teleoperated/telerobotic situations.

First, a delay compensation aid based on a model-free predictor was developed and implemented on a high fidelity driving simulation. Then, a human-subject experiment was conducted. Participants teleoperated an unmanned ground vehicle using a gaming steering wheel in a driving simulation with and without the delay compensation aid while performing a secondary task (auditory 1-back memory task). Driver's workload was measured by subjective rating, physiological measurement and secondary task performance.

This work has been written as a conference paper and accepted in Proceedings of the 13th ACM/IEEE International Conference on Human-Robot Interaction (HRI), 2018.

Lu, Shihan, Meng Yuan Zhang, Tulga Ersal, X. Jessie Yang. "Effects of a Delay Compensation Aid on Teleoperation of Unmanned Ground Vehicles." Proceedings of the 13th ACM/IEEE International Conference on Human-Robot Interaction (HRI), 2018. 

I also presented my work at 12th Graduate Symposium, UMich. The poster is here.

In this project, I explored the modeling of probabilistic trust inference for human-robot interaction. Hidden states related to human-robot interaction like trust, emotion and confidence are significant factors which can determine the strategies and dynamics of interactions.

However, to infer the trust between human and robot interactions is not simple, because

• States of trust are always unobservable and imperceptible, and even if they are reported, can they be easily mislabeled;

• Data of trust is often sparse, especially in time-critical situations;

• High correlation with other states increases the difficult of modeling trust in finite space.

Facing these problems, I am now using the Dynamics Bayesian Network (DBN) to infer human’s time-series latent trust states, based on the history of observed human behaviors. DBN is an effective way to handle time-series data and forecast next steps. By changing the structure of graphical models and tuning the weight of different observations, better prediction of human’s trust can be achieved. Also, this DBN model is being validated by the dataset from a navigation and detection task.