[Paper Review] Exploring Model Predictive Control to Generate Optimal Control Policies for HRI Dynamical Systems

Steven Jens Jorgensen, Orion Campbell+4 more|arXiv (Cornell University)|Jan 13, 2017

Social Robot Interaction and HRI17 references3 citations

TL;DR

This paper proposes using Mixed-Integer Model Predictive Control (MPC) to generate human-aware control policies for human-robot interaction (HRI) by modeling HRI scenarios as linear dynamical systems with cost functions for productivity, workload, connection, and awkwardness. The method successfully generated intended behaviors in simulation and on a humanoid robot, increasing eye contact by 25% when maximizing connection, though no significant difference in perceived connection was found in human trials.

ABSTRACT

We model Human-Robot-Interaction (HRI) scenarios as linear dynamical systems and use Model Predictive Control (MPC) with mixed integer constraints to generate human-aware control policies. We motivate the approach by presenting two scenarios. The first involves an assistive robot that aims to maximize productivity while minimizing the human's workload, and the second involves a listening humanoid robot that manages its eye contact behavior to maximize "connection" and minimize social "awkwardness" with the human during the interaction. Our simulation results show that the robot generates useful behaviors as it finds control policies to minimize the specified cost function. Further, we implement the second scenario on a humanoid robot and test the eye contact scenario with 48 human participants to demonstrate and evaluate the desired controller behavior. The humanoid generated 25% more eye contact when it was told to maximize connection over when it was told to maximize awkwardness. However, despite showing the desired behavior, there was no statistical difference between the participant's perceived connection with the humanoid.

Motivation & Objective

• To model human-robot interaction (HRI) scenarios as linear dynamical systems amenable to control-theoretic optimization.
• To develop a cognitive feedback controller using Model Predictive Control (MPC) with mixed-integer constraints to generate human-aware behaviors.
• To evaluate whether MPC can produce optimal control policies that balance productivity, workload, connection, and social awkwardness in HRI.
• To validate the approach in simulation and on a real humanoid robot through human-in-the-loop experiments.
• To investigate whether robot behaviors generated via MPC can meaningfully influence human perceptions of connection and comfort.

Proposed method

• Formulate HRI scenarios as linear dynamical systems with state variables representing human workload, battery level, eye contact, and connection/awkwardness metrics.
• Define a cost function that penalizes high workload, low productivity, awkwardness, or low connection, depending on the scenario.
• Implement Mixed-Integer MPC to solve the optimization problem under real-valued and binary constraints (e.g., robot can only move if battery is sufficient).
• Use a convex optimization framework (CVXPY with Gurobi) to solve the MPC problem in real time with a 3Hz control loop.
• Deploy the controller on a humanoid robot using the Wizard-of-Oz technique to simulate eye contact detection via manual input.
• Tune the MPC model based on intuitive parameter settings derived from Social Cognitive Theory and behavioral dynamics.

Experimental results

Research questions

• RQ1 Can HRI scenarios be effectively modeled as linear dynamical systems suitable for optimal control?
• RQ2 Can Mixed-Integer MPC generate control policies that balance competing objectives like productivity and human workload in assistive HRI?
• RQ3 Can MPC generate socially adaptive behaviors such as eye contact that maximize connection and minimize awkwardness?
• RQ4 Does the robot’s MPC-generated behavior significantly alter human perceptions of connection, comfort, or interest in social HRI?
• RQ5 How do implementation factors (e.g., hardware limitations, control loop lag) affect the perceived effectiveness of MPC-generated HRI behaviors?

Key findings

• The MPC controller successfully increased eye contact by approximately 25% when instructed to maximize connection compared to maximizing awkwardness.
• Despite the increase in eye contact, no statistically significant difference was found in participants’ perceived connection, comfort, enjoyment, or interest between the two robot behavior targets.
• Correlation analysis showed very low R² values (ranging from 0.002 to 0.012) between robot behavior and survey responses, indicating minimal relationship.
• Participants’ written feedback was more negative when the robot was programmed to maximize awkwardness, suggesting a qualitative perception difference despite no statistical significance.
• The humanoid robot’s eyelid motors failed during testing, which may have contributed to the lack of perceived effect, indicating hardware limitations impacted results.
• Many participants expected the robot to use backchanneling (e.g., head nods, voice cues), indicating that the current eye contact model alone is insufficient for deep social connection.