Dogs Obey Commands Given by Social Robots

As far as dogs are concerned, social robots have some humanlike authority

5 min read

Dog obeys commands from a robot
Photo: Social Robotics Lab/Yale University

One of the things that sets robots apart from intermittently animated objects like toasters is that humans generally see robots as agents. That is, when we look at a robot, and especially a social robot (a robot designed for human interaction), we tend to ascribe some amount of independent action to them, along with motivation at varying levels of abstraction. Robots, in other words, have agency in a way that toasters just don’t. 

Agency is something that designers of robots intended for human interaction can, to some extent, exploit to make the robots more effective. But humans aren’t the only species that robots interact with. At the ACM/IEEE International Conference on Human-Robot Interaction (HRI 2020), researchers at Yale University’s Social Robotics Lab led by Brian Scassellati presented a paper taking the first step toward determining whether dogs, which are incredibly good at understanding social behaviors in humans, see human-ish robots as agents—or more specifically, whether dogs see robots more like humans (which they obey), or more like speaker systems (which they don’t).

The background research on dog-robot interaction that forms the basis for this work is incredibly interesting. The paper is absolutely worth reading in its entirety, but here are a few nuggets of prior work that should help you understand how dogs interact with nonhuman animated objects:

Pongrácz et. al tested whether dogs followed commands from their guardians with various levels of embodiment. The guardians may be present in the same room as the dogs (i.e., 3D condition), or interacted with the dogs via live-stream life-size interactive videos (i.e., 2D condition), or interacted with the dogs with only their voices came out of a loudspeaker (i.e., 0D condition). Dogs followed the commands most reliably in the 3D condition. They followed the commands least consistently in the 0D condition, and their performances were between 3D and 0D condition in the 2D condition.

Lakatos et. al conducted a study to test how dogs responded to the pointing cues given by a PeopleBot with customized arms. The PeopleBot either exhibited human-like behaviors or no social behaviors, depending on the condition. A dog participant observed the robot interacting with the guardian either socially or mechanically for six minutes in the interaction phase. The robot then delivered a food reward for the dog. In the subsequent testing phase, the robot pointed to one of the two buckets with hidden food rewards. In the testing phase, dogs performed better in the condition with a social robot than with a nonsocial robot. However, no evidence suggested the mean performance with the social robot was significantly different from 50 percent, which is the chance level in two-choice tasks. Therefore, the dogs did not consistently follow the pointing cues provided by the social robot, even though dogs in general follow human pointing cues well.

To summarize, dogs don’t respond very well to commands from loudspeakers or video systems, and they also don’t really pay attention when a mechanical-looking robot points at things, even though dogs understand what pointing means when humans do it.

Curiously, Aibo (a doglike robot) tends not to be perceived by real dogs as a competitor for food, and dogs in general don’t interact with Aibo in doglike ways. Dogs often react to Aibo in other ways, but it’s more like “what the heck is that thing” rather than “that’s a weird dog,” similarly to how some dogs react to things like vacuums (robot or otherwise). So if dogs understand on some level that robot dogs aren’t actually dogs, and don’t interact with robot dogs in doglike ways,  how would dogs interact with social robots that are designed to interact with humans and therefore have some humanlike features? 

Results of the experiments showed that the dogs paid significantly more attention to the robot than the speaker, and were significantly more likely to follow a sit command from the robot

The Yale researchers put together an experiment that compared how dogs respond to commands given by a Nao to how dogs respond to the same commands given by a speaker system. A group of 34 dogs participated in the experiment, and each dog was tested with either the speaker or the Nao (but not both) in a room that also included a researcher and the dog’s guardian. After a brief intro to the testing environment, the robot or speaker called the dog’s name (using the same voice), and the researchers noted whether the dog paid attention at all. Then the robot or speaker would talk to the guardian for a bit in an attempt to “denovelty” itself, provide a treat to the dog, and then give a “sit” command, which was the real test.

Results of the experiments showed that the dogs paid significantly more attention to the robot than the speaker, and were significantly more likely to follow a sit command from the robot. Dogs obeyed the sit command over 60 percent of the time when it came from the robot, but less than 20 percent of the time when it came from the speaker, even if they did look a bit confused about the whole thing at times. 

While these results are certainly interesting, it’s important to emphasize that the goal here was, according to the researchers, to “answer the question of whether dogs could respond to a social robot at all.” The researchers weren’t (yet) trying to determine what factors might increase or decrease that likelihood, but instead, they were giving the dogs a sort of ideal opportunity. For example, the dogs’ guardians were instructed to interact with the robot, talking to it and making eye contact, to help encourage the dogs to see the robot as friendly and alleviate any potential anxiety while also drawing attention to the robot. 

It’s also not entirely clear exactly what the dogs are responding to. Lead researcher Meiying Qin highlighted some differences for us between a robot and a speaker that could have caused the behavioral differences, including:

  • A robot may be perceived as an agent, but not the loudspeaker;
  • A robot is embodied, while the speaker is not (the agent who provided the commands via the loudspeaker is not physically present in the testing room);
  • A robot provided both visual and audio cues, while a speaker provided only audio cues.

We asked Qin whether she thought it would make a difference if the robot was more or less humanoid, how much of a face it had, whether it smelled like anything, and other traits that dogs might associate with human-ness. “Since dogs are very sensitive to human social cues, the robot being a humanoid or not may make a difference,” Qin says. “However, if a non-humanoid robot behaved like an agent (e.g., behaved like a dog, or exhibit any social behaviors), dogs may also respond in a social manner.”

She explained that, in terms of whether the robot has eyes or not, or smells like a person, these factors could also impact how dogs respond to the robot. But Qin adds that the researchers need further evidence to give a more affirmative answer. “Whether the robot moves or not could affect the dogs differently,” she says. “A robot that just stands still without any movement may not present itself as an agent to the dog, and the dogs may not respond to such a robot socially. On the other hand, a robot that moves too much (e.g., the robot walks) or moves too fast will simply scare the dogs.”

Now that we have evidence that dogs do respond to social robots, the next step is to figure some of this stuff out. And it’s not just about making more effective social robots for dogs, of course—the larger context is that by studying how dogs behave toward social robots relative to humans, it can help us understand how social robots affect our own behavior, too.

“Dog Sit! Domestic Dogs (Canis familiaris) Follow a Robot’s Sit Commands,” by Meiying Qin, Yiyun Huang, Ellen Stumph, Laurie Santos, and Brian Scassellati from Yale University, was presented virtually at HRI 2020. You can read the full paper here.

The Conversation (0)