CtrlK
A social robot is an autonomous or semi-autonomous robot designed to interact and communicate with humans (and sometimes with other robots) by following social behaviors and rules attached to its role. Social robots form a distinct subset of robotics and a central research area within human-robot interaction (HRI). Unlike industrial robots, which prize precision and throughput, social robots are evaluated on whether people perceive their behavior as appropriate, engaging, comprehensible, and emotionally satisfying.
The field grew out of late-1990s work at the MIT Artificial Intelligence Laboratory, most famously Cynthia Breazeal's expressive head Kismet, and has since produced therapy robots used in dementia care, classroom assistants for children with autism, animatronic pets, telepresence avatars, museum receptionists, and a long line of consumer companions. Many of those consumer products have failed commercially despite enthusiastic launches; the field has churned through a graveyard of well-funded shutdowns including Jibo, Anki's Cozmo and Vector, Mayfield Robotics' Kuri, and Embodied's Moxie. The arrival of capable large language models after 2022 has revived the category by making fluent open-ended conversation feasible without scripting every dialogue branch by hand.
The most widely cited working definition was given by Christoph Bartneck and Jodi Forlizzi in 2004: a social robot is an autonomous or semi-autonomous robot that interacts and communicates with humans by following the behavioral norms expected by the people with whom it is intended to interact. Their framework characterizes social robots along five properties: form, modality, social norms, autonomy, and interactivity. A robot is more or less social depending on where it sits on each of those axes, not on whether it crosses a fixed threshold.
Breazeal's earlier work proposed a complementary taxonomy that distinguished four classes by how socially capable a robot really is. Socially evocative robots simply rely on people's natural tendency to attribute feelings to anything with eyes; they do not actually model the user. Social interface robots use human-like cues such as gaze and gesture so that interaction feels natural, but the underlying model of the human is shallow. Socially receptive robots respond to social cues from humans, learning passively from interaction. Sociable robots actively pursue social goals of their own, modeling the human as a partner with beliefs, desires, and intentions. Most products on the market today fall into the first two classes; almost none reach the fourth.
Social robots differ from voice assistants like Alexa and Google Assistant in that the body matters. Embodiment lets a social robot orient toward speakers, follow a person around a room, share gaze with an object on a table, hand things over, and use posture and proxemics. Studies in HRI have repeatedly shown that physical embodiment increases trust, persuasion, and learning outcomes compared with the same software running on a screen.
Social robots are usually grouped by application domain rather than by mechanical form. The boundaries between categories are loose, and many products straddle two or three.
| Category | Purpose | Representative robots |
|---|---|---|
| Companion robots | Daily emotional engagement, conversation, presence | Jibo, ElliQ, Buddy, Lovot, Samsung Ballie |
| Robot pets | Animal-like affection without animal-care burden | Sony Aibo, PARO, NeCoRo, Joy For All cats |
| Therapeutic robots | Clinical or quasi-clinical use in care settings | PARO, NAO, Milo, Stevie |
| Educational robots | Tutoring, language learning, coding instruction | NAO, Pepper, Cozmo, Vector, Moxie |
| Service and reception robots | Greeting, wayfinding, light retail tasks | Pepper, Wakamaru, Hitchhike, Reachy |
| Telepresence robots | Embodied remote presence for work, school, care | Beam, Double, Ohmni, BeamPro |
| Research platforms | Open hardware for HRI experiments | NAO, iCub, Furhat, Reachy, Pepper |
Most of these robots are stationary or wheeled. Bipedal social robots remain rare because the engineering cost of stable walking is hard to justify when the robot's job is to chat or to sit on a side table.
Before there was a recognized field, scattered projects hinted at what was coming. Toshitada Doi and colleagues at Sony explored entertainment robots in the late 1980s. Rodney Brooks's behavior-based architecture at the MIT AI Lab produced Cog in the early 1990s, an upper-torso humanoid intended to study social and developmental learning. The Furby toy from Tiger Electronics in 1998, while not a research robot, sold tens of millions of units by exploiting exactly the social-evocative response Breazeal would later formalize. Furby's commercial success made it clear that consumers were willing to pay for an object that pretended to have feelings.
Kismet was built by Cynthia Breazeal at the MIT Artificial Intelligence Laboratory beginning in 1997 and is widely credited as the first social robot in the modern sense. Kismet was a disembodied head with movable lips, ears, eyebrows, and eyelids. It used cameras and microphones to track faces and recognize prosody, and it could express anger, calm, disgust, fear, happiness, interest, sadness, surprise, and tiredness through its face and through nonsense vocalizations. Crucially, Kismet was designed as an infantile creature so that adults would naturally treat it the way they treat a baby, scaffolding the interaction without realizing they were doing the work. Breazeal's 2002 book "Designing Sociable Robots" laid out the theoretical framework, the architecture, and the empirical results from Kismet and gave the subfield its name.
Sony released the Aibo ERS-110 on 1 June 1999 at a retail price of about US$2,500. The first batch of 3,000 units sold out in twenty minutes online. Five generations followed through 2006, with the ERS-7 of 2003 being the most refined; total sales across all first-era generations exceeded 150,000 units. Sony killed the line in January 2006 as part of a profitability drive. Around the same period Omron released NeCoRo, a furry robotic cat marketed for emotional companionship.
The most influential therapy robot from this era was PARO, a baby harp seal designed by Takanori Shibata of Japan's National Institute of Advanced Industrial Science and Technology (AIST). Shibata began the project in 1993 and modeled the seal on harp seal pups he had observed in northeastern Canada, recording their cries to use for the robot's voice. PARO was first publicly exhibited in late 2001 and has been sold commercially since 2004 at around US$5,000 per unit. Guinness World Records named PARO the world's most therapeutic robot in 2002. PARO is used in dementia care, autism therapy, and pediatric pain management; clinical trials have shown reductions in stress hormones and decreased need for sedating medication.
Mitsubishi Heavy Industries launched Wakamaru in Japan in 2005, a 1-meter yellow humanoid designed by Toshiyuki Kita to live with elderly users, recognize family faces, read out email, and offer reminders. It cost roughly US$14,000, sold poorly, and was effectively retired into museum exhibits. The French company Aldebaran Robotics released NAO in 2008, a 58-cm bipedal humanoid that became the standard research platform for HRI labs and is now used in over 600 universities worldwide. SoftBank acquired Aldebaran in 2012 and in 2014 introduced Pepper, a 1.2-meter wheeled humanoid pitched as the world's first robot designed to read human emotion. Pepper saw significant deployment in Japanese banks, retail stores, and reception desks, although by the early 2020s SoftBank had paused production amid lukewarm demand.
A cluster of well-funded startups attempted to build the social-robot equivalent of the iPhone in the mid-2010s. Cynthia Breazeal founded Jibo in 2012 and ran a 2014 Indiegogo campaign that raised more than US$3.5 million, which at the time was the largest crowdfunding total for a tech product. Jibo eventually raised roughly US$73 million in venture funding. The robot, a 28-cm rotating tabletop unit with a single animated eye, finally shipped to backers in late 2017 (three years after the campaign), by which point Amazon Echo and Google Home had taken the voice-assistant market. Jibo Inc. closed in November 2018, sold its IP to a New York investment firm, and most servers were shut down in March 2019. Jibo broke the news of its own death to owners with a short farewell speech that briefly went viral.
Mayfield Robotics, a 2015 spinoff of Bosch's Startup Platform, built Kuri, a 50-cm wheeled home robot with expressive eyes and ambient sensing. Kuri shipped to backers in late 2017 at US$700. In July 2018 Bosch concluded there was no business fit, ceased manufacturing, and refunded all pre-orders. The company shut entirely by October 2018.
Anki, a San Francisco company founded by Carnegie Mellon graduates, raised US$200 million across multiple rounds and shipped Cozmo (2016) and Vector (2018), both palm-sized expressive robots clearly inspired by Pixar's Wall-E. Anki sold more than 1.5 million Cozmo units. A late-stage funding round collapsed and Anki shut down in April 2019. In December 2019 the IP was acquired by Pittsburgh-based edtech startup Digital Dream Labs, which later released Cozmo 2.0 and Vector 2.0 with limited success.
A contemporaneous IEEE Spectrum analysis attributed the wave of failures to a common pattern: hardware margins too thin to support cloud services, voice-assistant competition that ate the basic-utility use case, and entertainment value that wore off after a few weeks of novelty.
While consumer companions were collapsing, robots aimed at care continued to gain ground because the buyers (hospitals, care homes, government aging agencies) judged value differently. Intuition Robotics of Tel Aviv launched ElliQ in 2017, a small swiveling head and screen designed for adults aging at home. The New York State Office for the Aging began deploying ElliQ at scale in 2022; published outcome data showed users engaging with the device more than twenty times per day and a self-reported 95% reduction in loneliness scores in the pilot cohort. ElliQ 3, announced at CES 2024, integrates large language model technology to handle open-ended conversation while keeping a goal-based supervisor that screens responses for vulnerable users.
Groove X, founded by ex-SoftBank engineer Kaname Hayashi, released Lovot in Japan in December 2018. Lovot is explicitly designed not to be useful: it has no fixed task list and instead asks to be picked up, follows owners, and reacts to touch with around fifty body sensors. Pricing began near US$3,000 plus a monthly service fee. Lovot has remained in production and expanded to Hong Kong and other markets.
In special education, the NAO robot has been used for autism therapy in dozens of clinical settings. RoboKind's Milo, a 60-cm humanoid with a screen on its chest, launched in 2014 specifically for children with autism spectrum disorder; the company reports more than 400 units in active classroom use, mostly in North America, and a clinically designed curriculum delivered jointly with the Callier Center at the University of Texas at Dallas.
The public release of ChatGPT by OpenAI in late 2022, followed by Claude, Gemini, and others, removed one of the biggest engineering bottlenecks in social robotics: scripting believable open-ended dialogue. A new wave of products either embedded LLMs or were rebuilt around them. Embodied Inc. launched Moxie in 2020 for children aged five to ten and updated it in 2024 with GPT-style conversational capability, marketing it especially to families with autistic children at US$799. Samsung revived its Ballie concept at CES 2024 as a softball-sized rolling robot with a built-in 1080p projector and LLM-driven conversational interface, targeting a 2025 release. Numerous smaller startups have released LLM-driven companion robots aimed at desk users.
The new wave has not avoided the old failure modes. Embodied announced in December 2024 that an expected funding round had collapsed and that the Moxie cloud servers would be shut down within days, bricking thousands of robots that families had bonded with. The shutdown drew unusually emotional press coverage; an open-source project called OpenMoxie was rapidly organized to keep the hardware functional offline. The episode renewed long-standing debates about whether selling cloud-dependent companion robots to children is ethically defensible.
| Domain | Typical use | Notable deployments |
|---|---|---|
| Eldercare and aging in place | Loneliness mitigation, cognitive stimulation, medication reminders | ElliQ in New York State, PARO in Danish and Japanese care homes, Lovot in Japan |
| Dementia care | Calming, reducing agitation, lowering antipsychotic use | PARO in nursing homes worldwide |
| Autism intervention | Social-skills practice, emotional recognition, predictable interaction partner | Milo (RoboKind), NAO (multiple research consortia), Kaspar (University of Hertfordshire) |
| K-12 education | Coding, language learning, STEAM | NAO, Cozmo, Vector, Pepper |
| Higher education and research | HRI, developmental robotics, embodied AI | NAO, iCub, Pepper, Reachy, Furhat |
| Hospital pediatrics | Distraction during procedures, emotional support | NAO, MEDi (Calgary trials) |
| Retail and reception | Greeting, wayfinding, simple Q&A | Pepper in SoftBank stores, Wakamaru historically |
| Hospitality | Hotel check-in, novelty | Pepper in hotels, the (now closed) Henn na Hotel in Japan |
| Workplace telepresence | Meeting attendance, lab oversight, remote teaching | Double Robotics Double 3, Beam |
| Astronaut companionship | Stress reduction in isolation studies | PARO in JAXA Mars-analog studies |
The table below lists widely studied or commercially significant social robots. Status reflects the status of the originating company's commercial offering, not whether individual hardware units still function.
| Robot | Maker | First released | Form | Status |
|---|---|---|---|---|
| Kismet | MIT AI Lab (Breazeal) | 1998 (research) | Expressive head | Retired to MIT Museum |
| Aibo (ERS-110) | Sony | 1999 | Robot dog | Discontinued 2006, revived 2018 |
| PARO | AIST (Shibata) | 2004 commercial | Baby harp seal | Active |
| Wakamaru | Mitsubishi Heavy Industries | 2005 | 1 m humanoid | Discontinued |
| NeCoRo | Omron | 2001 | Robot cat | Discontinued |
| NAO | Aldebaran / SoftBank | 2008 | 58 cm bipedal humanoid | Active (research) |
| Beam | Suitable Technologies | 2012 | Telepresence | Acquired by Blue Ocean Robotics, sold by Awabot |
| Pepper | Aldebaran / SoftBank | 2014 | 1.2 m wheeled humanoid | Production paused 2021 |
| Buddy | Blue Frog Robotics | 2015 (crowdfunding) | Wheeled tabletop | Repeatedly delayed, intermittent shipments |
| Jibo | Jibo Inc. (Breazeal) | 2017 | Rotating tabletop | Discontinued 2018, servers down 2019 |
| Kuri | Mayfield Robotics (Bosch) | 2017 | Wheeled home robot | Discontinued 2018 |
| Cozmo | Anki | 2016 | Tracked tabletop | Anki shut 2019, IP at Digital Dream Labs |
| Vector | Anki | 2018 | Tracked tabletop | Anki shut 2019, IP at Digital Dream Labs |
| Lovot | Groove X | 2018 | Wheeled fuzzy companion | Active |
| ElliQ | Intuition Robotics | 2017, ElliQ 3 in 2024 | Tabletop head plus tablet | Active |
| Aibo (ERS-1000) | Sony | 2018 | Robot dog | Active |
| Milo | RoboKind | 2014 | 60 cm humanoid | Active (education) |
| Moxie | Embodied Inc. | 2020 | 38 cm tabletop humanoid | Discontinued December 2024 |
| Ballie | Samsung | CES 2024 prototype | Rolling sphere with projector | Announced for 2025 |
| Reachy | Pollen Robotics | 2020 | Open-source humanoid torso | Active |
| iCub | Italian Institute of Technology | 2008 | Child-sized humanoid | Active (research) |
| Furhat | Furhat Robotics | 2014 | Back-projected face on a torso | Active |
Anthropomorphism is the human tendency to attribute mental states and human-like traits to non-human entities. Social robotics depends on it. Designers deliberately add or remove cues (eyes, eyebrows, head tilt, vocal prosody, hesitation pauses, breath-like idle motion) to push users into treating the device as a social agent. Research by Bartneck, Eyssel, and others has shown that even minimal cues such as a pair of camera-lens eyes are enough to trigger the social response. Over-anthropomorphizing carries cost, however; users hold an obviously social-looking robot to higher conversational and social standards than they would hold a tabletop speaker.
The Japanese roboticist Masahiro Mori proposed in 1970 that as a robot is made more humanlike, the human emotional response becomes more positive and empathetic, until at a point very close to (but not quite at) human appearance, the response flips into revulsion. Mori called the dip the bukimi no tani genshou or uncanny valley. Movement amplifies the effect. Most successful social robots avoid the valley by stopping well short of realism: PARO is a seal, Lovot is fuzzy and abstract, Cozmo is a small box, Jibo had a single eye. Robots that have tried to look photoreal, such as Hanson Robotics' Sophia, are often cited as walking the edge of the valley.
Daniel Dennett's notion of the intentional stance, the idea that observers will treat a system as if it has beliefs and desires whenever doing so usefully predicts its behavior, underlies many design choices in social robotics. If the robot's behavior is consistent enough that the user can adopt the intentional stance, the interaction works even when no real understanding sits underneath. Critics including Sherry Turkle have warned that this can produce relationships that feel real to the user but are not reciprocated, with implications for vulnerable populations such as children and people with dementia.
A recurring HRI finding, summarized in meta-analyses by Wainer, Feil-Seifer, Shell, and Mataric, is that the same software produces stronger persuasion, learning, and engagement when delivered through a physical robot than through a screen avatar or voice-only assistant. The effect size shrinks with task complexity but persists across studies.
| Researcher | Affiliation | Contribution |
|---|---|---|
| Cynthia Breazeal | MIT Media Lab; founder of Jibo | Pioneered sociable robotics with Kismet, authored "Designing Sociable Robots" (2002) |
| Takanori Shibata | AIST, Japan | Designer of PARO, leading figure in robot-assisted therapy |
| Masahiro Mori | Tokyo Institute of Technology | Proposed the uncanny valley hypothesis (1970) |
| Maja Mataric | University of Southern California | Socially assistive robotics, founder of the field; stroke rehabilitation and autism work |
| Brian Scassellati | Yale University | Social robotics for autism intervention, joint attention |
| Kerstin Dautenhahn | University of Waterloo (formerly Hertfordshire) | Kaspar robot for autism, theoretical foundations of HRI |
| Hiroshi Ishiguro | Osaka University | Geminoid androids, Telenoid, theoretical work on presence and the uncanny valley |
| Christoph Bartneck | University of Canterbury | Foundational definition of social robots, godspeed questionnaire for HRI evaluation |
| Selma Sabanovic | Indiana University Bloomington | Cultural and ethnographic studies of robots in care |
| Bilge Mutlu | University of Wisconsin-Madison | Designing robot gaze, gesture, and conversational behavior |
| Sherry Turkle | MIT (sociology) | Critical scholarship on emotional bonds with robots and AI |
The consumer social-robot category has a high failure rate. Common patterns across the canceled products include:
Care-sector robots have had a higher survival rate because purchasers (clinics, care homes, government programs) make purchase decisions on measured outcomes such as reductions in agitation, antipsychotic use, or loneliness scores rather than on novelty.
Social robots raise distinctive ethical questions because their entire design goal is to be treated as a social agent by people who know they are not. Key debates include:
Groups including the IEEE Global Initiative on Ethics of Autonomous and Intelligent Systems and the Foundation for Responsible Robotics have published guidelines on these issues, though regulation specific to social robots remains rare.
As of 2026 the field shows several active trends: