Social and assistive robots

Not every robot that works with people is there to assemble parts. Some are designed to engage people socially: to tutor a child, keep an older adult company, coach someone through rehabilitation exercises, greet hotel guests or guide visitors in a museum. Others assist people physically with daily living: robotic arms mounted on wheelchairs, feeding devices, exoskeletons and smart walkers. This lesson looks at both families, at the psychology that makes people respond to robots as social beings, at the norms of personal space that robots must respect, at the famous and much-debated uncanny valley hypothesis, and at the ethical questions raised when machines are built to be liked.

Kinds of social and assistive robots

David Feil-Seifer and Maja Matarić defined socially assistive robotics in 2005 as the intersection of assistive robotics, which helps people with disabilities or illness, and socially interactive robotics, which communicates through social channels. A socially assistive robot helps through social interaction rather than physical contact: it motivates, reminds, coaches and encourages. Physically assistive robots, by contrast, manipulate objects or support the body. Some systems combine both, such as a rehabilitation robot that moves a patient's arm and also encourages them with speech and facial expressions.

Paro, a robotic baby harp seal developed in Japan by Takanori Shibata, is one of the best-known examples. It responds to touch, sound and light with movements and sounds, and was cleared by the United States Food and Drug Administration as a class II medical device in 2009, a regulatory classification rather than proof of effectiveness. Studies in long-term care have reported calming effects for some people with dementia, but systematic reviews generally rate the evidence as limited by small samples, short durations and difficulty blinding. This pattern recurs throughout social robotics: encouraging early results, weaker evidence once larger and longer studies are run, and strong dependence on context.

Why people treat robots as social

People readily respond to machines as if they were social actors. In the 1990s, Byron Reeves and Clifford Nass described the media equation: in experiments, people applied social rules such as politeness and reciprocity to computers, even when they knew they were dealing with a machine. This became known as the computers are social actors, or CASA, paradigm. Robots, with bodies that move in shared space, trigger these responses even more strongly. People name their vacuum robots, apologize when they bump into a delivery robot and feel uncomfortable when a robot is kicked in a video.

Nicholas Epley and colleagues proposed in 2007 that people anthropomorphize, attributing humanlike minds to non-human agents, more when three factors are present: when humanlike knowledge is readily accessible to apply (elicited agent knowledge), when they are motivated to understand and predict the agent's behaviour (effectance motivation), and when they want social connection (sociality motivation). This helps explain why lonely people may anthropomorphize more strongly, and why robots that behave unpredictably invite people to attribute intentions to them. Designers can encourage or dampen anthropomorphism through appearance, voice, movement and naming, and should do so deliberately, since raising social expectations that the robot cannot meet leads to disappointment.

Personal space and social navigation

Edward Hall's work on proxemics in the 1960s described how people use distance in social interaction. For the North American middle-class adults he studied, he proposed four zones: intimate distance up to about 0.45 m, personal distance from about 0.45 to 1.2 m, social distance from about 1.2 to 3.6 m and public distance beyond about 3.6 m. The exact values vary with culture, relationship, age, gender and setting, but the idea that people expect others to respect zones of space applies to robots as well. Studies have found that people allow robots closer when the robot approaches from the front rather than from behind, when it moves more slowly, and when they have more experience with robots, and that people keep a greater distance from a robot that looks directly at them.

Social navigation applies these ideas to mobile robots moving among people. Rather than treating people as moving obstacles to avoid at a minimum safe distance, a socially aware planner adds costs for passing too close, crossing in front of someone, cutting between people who are talking, or approaching from behind. It follows local conventions, such as keeping to one side of a corridor, and signals its intentions early, as described in lesson 3. Learned methods trained on pedestrian data are increasingly used, but they can inherit the conventions of the place where the data was collected.

# Classifying approach distances into Hall's zones (synthetic trial data)
import numpy as np
rng = np.random.default_rng(430)
front = rng.normal(1.05, 0.25, 40)      # stopping distance chosen when the robot approaches from the front, m
behind = rng.normal(1.45, 0.30, 40)     # approaching from behind, m
edges = [0.0, 0.45, 1.2, 3.6, np.inf]
names = ["intimate", "personal", "social", "public"]
for label_, d in (("front", front), ("behind", behind)):
    counts = np.histogram(d, bins=edges)[0]
    shares = ", ".join(f"{n} {c / d.size:.0%}" for n, c in zip(names, counts))
    print(f"{label_:6s} mean {d.mean():.2f} m | {shares}")
print(f"difference in means: {behind.mean() - front.mean():.2f} m")

front  mean 1.06 m | intimate 2%, personal 72%, social 25%, public 0%
behind mean 1.52 m | intimate 2%, personal 15%, social 82%, public 0%
difference in means: 0.46 m

The uncanny valley

In 1970 the Japanese roboticist Masahiro Mori published a short essay proposing that as a robot becomes more humanlike, people's affinity for it rises, until it becomes almost but not quite human, at which point affinity drops sharply into what he called bukimi no tani, translated as the uncanny valley. Affinity then rises again for a fully healthy human. Mori suggested that movement deepens the effect, and that designers should aim for the first peak, a clearly non-human but appealing design, rather than risk the valley. An authorized English translation by Karl MacDorman and Norri Kageki appeared in 2012. The essay offered no data; it was a thought piece.

The uncanny valley became hugely influential in robotics, animation and game design, but the evidence for it is contested. Some studies find valley-like dips; for example, Maya Mathur and David Reichling in 2016 rated 80 real robot faces and found a dip in likeability for faces that were highly but not perfectly humanlike. Other studies find no valley, find that the effect depends heavily on how human likeness and affinity are measured, or explain the discomfort by other factors. Critics raise several points.

For designers, the practical lesson is narrower than the popular version. Mismatched cues, such as a realistic face with jerky motion or a humanlike voice from a machine-like body, often produce discomfort, and raising social expectations through humanlike appearance sets up disappointment when the robot cannot meet them. Many successful social robots therefore use stylized, clearly robotic or animal-like designs with expressive but simple features. That is a sensible heuristic supported by experience, not proof that a universal valley exists.

# Is there a valley? Comparing a straight line with a cubic fit on synthetic ratings
import numpy as np
rng = np.random.default_rng(430)
x = rng.uniform(0, 1, 120)                         # human-likeness score
y = 2 + 3 * x + rng.normal(0, 1.2, x.size)         # true relationship has no valley
def aic(deg):
    c = np.polyfit(x, y, deg); r = y - np.polyval(c, x)
    n, k = x.size, deg + 2
    return n * np.log(np.mean(r**2)) + 2 * k, c
for deg in (1, 3):
    a, c = aic(deg)
    print(f"degree {deg}: AIC {a:7.1f}, coefficients {np.round(c, 2)}")
grid = np.linspace(0, 1, 101); cub = np.polyval(aic(3)[1], grid)
print(f"cubic fit has a local dip: {bool(np.any(np.diff(np.sign(np.diff(cub))) > 0))}")

degree 1: AIC    43.5, coefficients [3.4  1.91]
degree 3: AIC    46.9, coefficients [ 4.06 -6.48  6.23  1.64]
cubic fit has a local dip: False

Long-term interaction and the novelty effect

Many social robot studies last a single session. In longer deployments, interaction often falls after the first days or weeks as the novelty wears off; a robot that children crowded around in the first week may be ignored by the fourth. Robots that sustain engagement tend to offer real usefulness, variety, memory of past interactions and adaptation to the person. Researchers model engagement over time and design studies that run long enough to see past the novelty peak. A common simple model treats daily engagement as a stable baseline plus a novelty component that decays exponentially.

Physically assistive robots

Physically assistive robots aim to increase independence for people with disabilities and older adults. Wheelchair-mounted robotic arms let users with limited upper-limb function eat, drink, open doors and pick up dropped objects. Feeding robots present spoonfuls of food. Exoskeletons support walking or reduce strain for workers lifting loads. Smart walkers and robotic beds assist mobility and transfers. The HRI challenges are considerable: control interfaces must work for users with limited motor function, often via joysticks, sip-and-puff switches, head arrays or shared autonomy (lesson 6); the robot operates close to the face and body, so safety is critical; and users want control over their own activities rather than having a robot decide for them. Cost, funding through public or private insurance, and training also determine whether a device is used. Lesson 10 returns to inclusive design and disability perspectives.

Ethics of social robots

Social robots raise particular ethical concerns. Amanda and Noel Sharkey and others have asked whether it is deceptive to give a robot behaviours that suggest feelings it does not have, especially for people with dementia or for young children who may not understand. Others argue that people enjoy fictions knowingly, as with pets and stories, and that the key question is whether the person benefits and keeps their dignity. Concerns about replacing human care are also prominent: a companion robot in a care home might reduce staffing pressure, but if it replaces visits or conversation with staff, residents could become lonelier. Robots with cameras and microphones in homes collect sensitive data (lesson 11), and children's interactions with robots raise questions about persuasion and attachment. Responsible design involves the people affected, family and care staff in decisions, and is honest about what the robot is and does.