Abstract
Robots start to play a role in our social landscape, and they are progressively becoming responsive, both physically and socially. It begs the question of how humans react to and interact with robots in a coordinated manner and what the neural underpinnings of such behavior are. This exploratory study aims to understand the differences in human-human and human-robot interactions at a behavioral level and from a neurophysiological perspective. For this purpose, we adapted a collaborative dynamical paradigm from Hwang et al. (1). All 16 participants held two corners of a tablet while collaboratively guiding a ball around a circular track either with another participant or a robot. In irregular intervals, the ball was perturbed outward creating an artificial error in the behavior, which required corrective measures to return to the circular track again. Concurrently, we recorded electroencephalography (EEG). In the behavioral data, we found an increased velocity and positional error of the ball from the track in the human-human condition vs. human-robot condition. For the EEG data, we computed event-related potentials. To explore the temporal and spatial differences in the two conditions, we used time-regression with overlap-control and corrected for multiple-comparisons using Threshold-Free-Cluster Enhancement. We found a significant difference between human and robot partners driven by significant clusters at fronto-central electrodes. The amplitudes were stronger with a robot partner, suggesting a different neural processing. All in all, our exploratory study suggests that coordinating with robots affects action monitoring related processing. In the investigated paradigm, human participants treat errors during human-robot interaction differently from those made during interactions with other humans.
Introduction
As robots transition from isolated industrial tools to social collaborators, understanding the "human-in-the-loop" becomes critical. While "second-person neuroscience" has extensively mapped how two humans coordinate, we know surprisingly little about the neural underpinnings of human-robot interaction (HRI) in real-time, dynamic settings. Most HRI studies rely on static stimuli or turn-based tasks. This paper addresses the gap in dynamic coordination—situations where a human and a robot must continuously adjust to each other's physical forces to achieve a shared goal. Specifically, the authors investigate action monitoring: how our brains detect and evaluate errors during joint tasks. Does the brain treat a robot’s "mistake" the same way it treats a human partner's slip-up?
Methods and Materials
The researchers adapted a collaborative dynamical paradigm where dyads (either two humans or one human and one robot) held opposite sides of a tablet. Their goal was to balance a virtual ball on a circular track while moving it counter-clockwise.
Experimental Setup
- The Robot: An ABB YuMi robot (non-humanoid) was used, controlled via a Cartesian space velocity controller.
- The Perturbation: To trigger action monitoring, the ball was randomly "pushed" off-track using a Poisson distribution (λ=4s), forcing the dyad to perform corrective maneuvers.
- Participants: The study involved 16 participants (data from 12 were used in the final analysis).
- Conditions: The experiment followed a 2x2 design: Partner (Human vs. Robot) and Sonification (Auditory feedback on/off).
Data Analysis
- Behavioral Data: High-frequency logs of ball position, velocity, and error were recorded.
- EEG Analysis: The team recorded 64-channel EEG. To handle the messy nature of continuous, dynamic movement, they employed a sophisticated linear deconvolution approach (using the unfold toolbox). This allowed them to control for overlapping neural responses and behavioral covariates like ball velocity and eccentricity. They used Threshold-Free Cluster Enhancement (TFCE) to correct for multiple comparisons across all electrodes and time points.
Results
Behavioral Findings: The Accuracy-Speed Trade-off
Humans were significantly faster when working with other humans (~265 deg/sec) but made more errors. With the robot, they were slower (~159 deg/sec) but much more precise. This suggests humans may adopt a more cautious, "monitoring-heavy" strategy when paired with a non-human agent. A significant main effect of "Partner" was found for both velocity (p < .0001) and mean error (p < .0001).
Neurophysiological Findings: Neural Sensitivity to Robot Errors
Despite the higher accuracy in the robot condition, the brain's Event-Related Potentials (ERPs) in response to perturbations were stronger with the robot partner. TFCE revealed two significant clusters for the "Partner" effect:
- Early Cluster (230–270ms): Peak amplitude of -2.8µV at electrode FC1.
- Late Cluster (515–605ms): Peak at -1.2µV at electrode FC2.
The differences were most pronounced at fronto-central electrodes between 200-300ms post-perturbation. This timing and location align with the Feedback-Related Negativity (FRN), a component associated with error detection and the evaluation of outcomes.
Discussion
This exploratory study reveals that the human brain processes collaborative errors differently when working with a robot versus another human. The study highlights that coordination isn't just about verbal cues; it's about "implicit sensorimotor communication." The brain appears to "tune" its monitoring system based on the perceived nature of the partner.
Limitations and Future Work
As an exploratory study, the sample size (N=12) is small, necessitating replication with larger cohorts. Furthermore, the YuMi robot is clearly non-humanoid; future research should investigate if more humanoid robots elicit neural responses closer to the human-human condition. The authors speculate that the stronger neural response with robots might be due to lower trust or a lack of a "mental model" for the robot's actions. Directly measuring trust or perceived agency in future EEG studies could clarify why the brain works harder to monitor a robot. Finally, future work could utilize "hyperscanning" (recording two brains simultaneously) to see how neural coupling changes when the partner is silicon-based.
Summary updated