A paper by Wonse Jo of Incheon National University outlines a controlled human–robot interaction experiment using Unitree’s U1 humanoid robot. The planned study would record several kinds of evidence at once: a participant’s physiological responses and movements, the robot’s motion and surroundings, and the participant’s own account of the interaction.
That matters because “engagement” is not a single sensor reading. A person may look toward a robot while feeling confused, or complete a task successfully while becoming increasingly overloaded. The proposed system is designed to compare these signals rather than treat any one of them as a complete answer.
The work is a research design, not a completed experiment or dataset. Its future work section describes implementing the protocol and collecting data later.
The experiment would change task difficulty, not just observe conversation
Jo’s protocol places participants in a controlled laboratory and asks them to collaborate with the robot on object-handling and movement tasks. Each of the planned 30 participants would complete three conditions, with the order randomized to reduce the effect of simply getting used to the procedure.
The three conditions are organized as task zones.
In the high-complexity zone, the robot would instruct the participant to retrieve five specified components, provide real-time validation and correction, and then have the participant arrange the components in a defined sequence. That creates frequent instruction-and-response exchanges.
The moderate condition would make the participant interpret descriptions rather than receive the target items’ names directly. Feedback would come only after the participant indicated that the task was complete.
The low-complexity condition would combine brief casual dialogue with clearly specified components, followed by immediate confirmation after each action.
This structure gives the researchers a practical comparison. If physiological responses, movements or self-reported workload change between zones, the team can examine whether those changes track the intended increase in complexity. It also gives the robot’s actions a timeline: an instruction begins, the participant moves, the robot responds, and the task ends.
One stream watches the person; another watches the machine
The proposed architecture divides the measurements into human-centric and robot-centric streams.
On the participant side, a wearable EmotiBit device would collect electrodermal activity, or EDA, and photoplethysmography, or PPG. EDA measures changes in skin conductance; PPG records an optical pulse signal. A wearable inertial measurement unit would also capture motion.
These measurements are not being used alone to declare that someone is engaged. They would sit alongside behavioral and subjective information. The protocol includes a 15-item questionnaire after each task zone, with every response rated on a five-point Likert scale. The questions cover areas including perceived cognitive demand, communication quality, responsiveness, social perception and overall satisfaction.
That self-report is important because physiological signals can show that something changed without explaining what the change meant. A participant might become more activated because a task is interesting, demanding or uncomfortable. The questionnaire provides the participant’s interpretation after the task.
The robot-side stream would record joint encoder data and robot-mounted inertial measurements to describe the U1’s movement. Three-dimensional LiDAR and RGB-D sensing would add spatial context, including information about positions and distances in the interaction area.
Together, the streams could connect a person’s response to a specific event. For example, a change in a participant’s motion or physiological signal could be compared with the robot beginning an instruction, moving its body, delivering feedback or completing a task.
Synchronization is the engineering core
The proposed study’s hardest practical problem is not collecting isolated sensor readings. It is lining them up.
A pulse signal, a wrist motion measurement, a robot joint angle and an RGB-D frame arrive in different formats and at different rates. The paper proposes using a unified acquisition clock and resampling the data onto a common timeline. The data streams would be synchronized through Robot Operating System 2, or ROS2.
The planned processing pipeline would retain both raw and preprocessed data. EDA and PPG would undergo filtering and outlier removal intended to reduce noise and motion artifacts. Human and robot inertial data would be adjusted for gravity, coordinate systems and units. Robot joint readings would be converted into angular-velocity profiles, with motion-onset timestamps extracted for segmentation.
The task zones themselves would be marked using protocol timestamps, while robot events such as instruction onset and task completion would provide additional markers. Metadata would describe sensor configurations, coordinate conventions and preprocessing settings.
For operators and researchers, this is the difference between a collection of recordings and a usable experiment. Without reliable timing, it becomes difficult to tell whether a physiological response followed a robot action, a participant’s own movement or a delay in the sensing system.
What the proposal could enable—and what it cannot show yet
If implemented as described, the study could let researchers compare three views of collaboration: what the participant did, what the participant’s body signaled and what the participant said afterward. Repeating the same design across three complexity levels could help identify patterns associated with workload, communication quality or interaction satisfaction.
The longer-term aim is to support models that estimate user state and eventually help robots adapt their behavior. That application remains future work, not a demonstrated capability of the U1 in this study.
The immediate consequence is more modest and more useful: the experiment defines a testable way to ask whether a person is coping with a collaborative task, rather than relying on a camera view or a post-task impression. The next meaningful step is implementation—recruiting the planned participants, collecting synchronized recordings and checking whether the signals agree with observed task performance and participant reports.