A novel decision-making architecture for human-robot collaboration: From mirror neurons to conflict-free interactions

Sobhani, Mehdi

Abstract

Inspired by the role of mirror neurons and the importance of predictions in joint action, a novel decision-making structure is proposed, designed and tested for both individual and dyadic action during real-world human-human and human-robot experiments. The structure comprises models representing individual decision policies, policy integration layer(s), and a negotiation layer. The latter is introduced to prevent and resolve conflicts among individuals through internal simulation rather than via explicit agent-agent communication.

As the main modelling tool, Dynamic Neural Fields (DNFs) were chosen. Data was captured from human-human experiments with a decision-making task performed by either one or two participants. The task involves choosing (picking) and placing blocks one by one from seven wooden blocks to create an alpha/numeric character on a kind of mechanical model of a 7-segment display. The task is designed to be as generic as possible. Recorded hand and blocks movements were used for developing DNF-based models by optimising parameters using a genetic algorithm.

Results show that decision policies can be modelled and integrated with acceptable accuracy for individual performances. In the dyadic experiment, using only individual models without the negotiation layer, the model failed to resolve conflicts. However, with the implementation of a negotiation layer, this problem could be overcome.

To Analyse the proposed model for a human-robot collaboration task, first, the role and efficacy of the negotiation layer of the architecture are assessed. Then, in a “Wizard of Oz” experiment, the performance of the complete architecture is compared with that of a human decision-maker. The same task of using wooden blocks to create characters in a 7-segment display is used in both experiments.

Results show a significant improvement in terms of the chosen objective and subjective measures when the robot uses the complete architecture with the negotiation layer. No significant difference was found for any of the measures between the human decision-maker and the complete model. Although the robot with a human decision-maker scored slightly better in all measures, a further Bayesian comparison of the data suggests a high probability of similarity between the model and the human decision-maker. This was further illustrated by a qualitative analysis of the post-experiment interview questions; in answering the third question, when asked which condition is more human-like, 17 participants identified that the robot using the complete model was like working with a human, and an equal number opted for identifying the robot controlled by a human decision-maker as being human-like. In addition, answering the first question, 6 participants found no difference between the robot being controlled by a human decision-maker and being controlled using the complete model.

The proposed decision-making structure based on DNFs is developed and tested for a simple pick-and-place task. However, the main primitive underlying action of this task, pick-and-place, is indeed part of many more complex tasks people perform in their day-to-day life. Paired with the possibility to gradually evolve the architecture by adding new policies on demand, the architecture provides a general framework for modelling decision-making in joint action tasks.

To demonstrate the generality of this ability, a car assembling task was used in a "Wizard of Oz" experiment. Similar to the previous experiment, participants worked with a robotic arm to perform the task. Each participant repeated the task 6 times, 3 times for each condition, Model or Wizard, in a random order. Again, no significant difference was found between the two conditions and the Bayesian comparison showed a high probability of similarity. When data were sorted based on the order of trials, a significant difference was found between the task completion time from the first trial to the last. This could be due to the fact that participants were repeating the same task, however, given the low number of participants for this experiment, which was executed on a small scale only to illustrate the potential for the ability to transfer the capability to a different task, further analysis is required in the future.

• Thesis Type: Thesis
• Deposit Date: Apr 17, 2023
• Publicly Available Date: Nov 21, 2023
• Public URL: https://uwe-repository.worktribe.com/output/10630076
• Award Date: Nov 21, 2023