HARK (as in, “Hark, who goes there?”) as an acronym may be a bit of a stretch:

Honda Research Institute Japan
Audition for
Robots with
Kyoto University

but the technology itself is something cool that could play a significant role in a not-too-distant future of education.


Wherefore art thou, HARK?

Yesterday I was able to join two faculty from UNC’s Physics department, two representatives from the Nomura Research Institute, and a final representative from Honda to learn about HARK.  HARK grew out of Honda’s robotics lab as a hardware designed to allow robots to process complex and multi-source audio within a positional framework.  That’s a bit much to process at once, so here’s a video demonstration.

The HARK audio device
The HARK audio device

The hardware is a dome-shaped object containing eight microphones that sits on a flat surface (preferably a flat surface perpendicular to the vector of gravity… of course), USB connectors, and on the device that I saw a printed band of alphabetically ordered letters surrounding the base.

The key features of the device are the ability to simultaneously record multiple sound inputs, differentiate and separate each of the inputs (over 90% reliability with 9 or fewer speakers) based on a fixed location, and work in the context of ambient or background noise.


Educational application

At Honda they have used the HARK device alongside of analytical software to evaluate group participation and communication patterns within teams and students in corporate settings.  HARK is able to measure total participation time, sequential conversation patterns, and interruptions.  As an example one concept they operationalize is “influence”, which is defined as a pattern of interactions occurring after one person has spoken.

Analysis of the results according to HARK’s algorithm identifies three types of group participants: dominators, facilitators, and idea creators.  These identities are conferred based on a few criteria including numbers of interruptions, total time talked, number of times talked (which could provide an average talking length), and patterns of interactions after a speaker has spoken.  One interesting graphic is that interruptions are measured as a set of person to person interactions on a grid.  Analysis of that grid can help to identify whether an individual equally interrupts everyone in the group, or only a single person.

The hope is that these analyses can operate with minimum work required by the faculty member, and operate at a scale to make it a benefit in a large class with multiple groups where a faculty or TA cannot always be present to actively view all interactions.

Within an educational setting there is a concept of an ideal grouping of three people: a manager, a skeptic, and a note-taker.  You can see this outlined in a discussion for doing group work in a different physics class at http://groups.physics.umn.edu/physed/Research/CGPS/FAQcps.html.  Conceptually this is an ideal arrangement of people for any type of group work, and is designed to produce optimal results.  While we were unable to draw clear parallels between this taxonomy and what HARK produces (dominator, facilitator, idea generator), the form seems similar, and it was suggested that knowing these traits about members in a group could help to intentionally create groups  that perform optimally.



A few items were noted as potential drawbacks with HARK.

  • When looking at how “influence” is operationalized it could be argued that this is not the best measure of influence, and it was additionally noted that all influence isn’t positive influence.
  • We brought up that the system is limited to only sound interpretation.  Just because one person is not speaking does not mean that they are not engaged, as body language could indicate if that student was quiet, but also leaning forward into the conversation and actively listening.
  • There were questions regarding the construct validity of the three types of individuals identified, and whether this is a substantial enough taxonomy to effectively build groups.  It would also be helpful to see data on the construct reliability of the HARK system – that is, whether it identifies students as accurately as a means tested group of faculty would.
  • Finally, at this point there seems to be a lot of work (some of it one-time) to plan for and make use of the HARK analysis.  These preparatory steps include orienting the HARK device at each table (preferably in a manner that could not be tampered with), labeling and arranging seats around a table and keeping them relatively fixed, ensuring that students don’t relocate during a session, and then effecting data driven changes based on the analysis.


Ultimately, the faculty present felt that the HARK system did not confer enough of an ROI to make the cost of implementation worthwhile.  They felt that they were not in a situation where the scale afforded would be necessary (they felt that they could already effectively observe the entire class), and that there was not yet enough data suggesting what optimal group assignment would be after class members were categorized.  I and the faculty thought that there were opportunities for additional research both to establish those “what next?” best practices, as well as methods to make incorporation of the hardware easier in a large class setting.

One structural advantage is that because a single device can effectively differentiate between nine students, then it would not be necessary to have a number of HARK devices equal to the number of groups based on the arrangement of those groups in a class.

My own contribution is that this is a technology worth revisiting as natural language processing improves to be able to actively monitor each audio thread.  Adequate growth in NLP/AI (Natural Language Processing/Artificail INtelarchitectures could help the system to be additionally effective in analyzing not only the interactions, but also the value of the content; it would know whether actively speaking members are actually speaking on topic, and additionally be able to isolate and surface for the faculty member themes across various groups.


Additional Information

I encourage you to check out the HARK page to see the other types of projects they’re working on.  The one that catches my eye right away is the HARK-Kinect (the Microsoft visual input device for the Xbox) item.

Comments are closed.