Robots are quietly becoming one of the most interesting accessibility technologies in the field.
They are not a replacement for software accessibility or inclusive design. They are a genuinely new category of collaborator — one that can navigate unfamiliar indoor spaces, physically guide someone’s hand, attend a class from another continent, teach a dance pose you can reach out and touch, or extend a performer’s body onto a stage.
Across ASSETS, W4A, CHI, HRI, and related venues, a coherent picture is emerging. Not the shiny humanoid future pop culture suggests. Something more tractable, more honest, and more human.
Here is what 36 peer-reviewed studies are telling us.
“All Robots Are Disabled”
Elmimouni, Šabanović, and Rode’s 2024 study of telepresence robots in university classrooms drew on Williams’ provocative argument that all robots are disabled. Their 22 students attending remotely via Beam robots ran into real limitations — 19 couldn’t adequately see slides, 11 struggled to hear, 13 lacked a sense of what was going on.
Telepresence users, the authors argue, become cyborgs — gaining new abilities (remote participation) while acquiring new disabilities (restricted vision, impaired hearing, limited mobility). One participant compared navigating narrow hallways with the robot to “exactly how it feels like to be in a wheelchair.”
The framing matters: solutions benefiting telepresence users also benefit students with disabilities. Universal design, with a robot in the room.
Navigation: The Robot as Guide
Navigation for blind travellers is where robotics research has gone deepest. A multi-year programme across Carnegie Mellon, IBM Research, and the Miraikan science museum in Tokyo — led by Chieko Asakawa and colleagues — has taken it from lab prototype to daily operational deployment.
CaBot (Guerreiro et al., 2019) was the starting point: a suitcase-shaped autonomous guide with LiDAR, vibro-tactile handle feedback, and bone-conduction audio that preserves environmental awareness. Ten blind participants completed every navigation task with zero errors and scored CaBot an average 88 on the System Usability Scale — comfortably in the “excellent” band. One participant: “I’ll be calling my dog CaBot later!”
BBeep (Kayukawa et al., 2019) reframed collision avoidance as a two-party problem. Pedestrians often don’t notice blind travellers — they’re on phones, facing away. An external speaker emitting tones audible to both user and crowd reduced imminent collision risks to 0.41 per encounter, versus 2.00 using a private bone-conducting headset. Participants judged this appropriate for airports and malls — but not hospitals or libraries. Social acceptance of an alert depends on the setting, not the sound.
BlindPilot (Kayukawa et al., 2020) tackled the last few metres — approaching a specific chair, counter, or door. Reorientation events dropped from nearly three per task with audio cues to effectively zero with a robot handle in hand. Global and local navigation demand different interfaces.
Miraikan (Kayukawa et al., 2023) was the first real-museum deployment. Eight totally blind participants explored a 2,100 m² floor during regular opening hours — all independently, no safety incidents. More striking still, 99.1% of 108 sighted bystanders agreed assistive robots should be introduced in museums. The question had shifted from whether autonomous assistive robots are socially acceptable to how the division of labour between machine navigation and human interpretation should be designed.
AI Suitcase field trials (Takagi et al., 2025) reports permanent daily operation at Miraikan, launched April 2024 — 1,288 users in its first five months. Eighty-three percent of visually-impaired respondents were satisfied; 71% felt safe. An outdoor trial returned a SUS score of 83.2. But the paper’s most striking finding is uncomfortable: some blind participants felt less safe when the suitcase made them less visibly identifiable as blind pedestrians — invoking Thomas Carroll’s “loss of obscurity” and Japan’s legal requirement that solo blind pedestrians use a white cane or guide dog.
Visibility is sometimes a safety feature.
WanderGuide (Kuribayashi et al., 2025) targeted wandering rather than wayfinding — map-less operation using SLAM and GPT-4o scene descriptions. But 28.6% of auto-mode scene descriptions contained errors, rising to 60% in conversation mode. Off-the-shelf vision-language models still under-specify named exhibits and shop identities. The weakest link in the pipeline is at the top of it.
Beyond Omakase (Kamikubo et al., 2025) reframes autonomy itself. Omakase is Japanese for “I leave it to you,” and the authors argue that current systems default to exactly that: the robot decides; the user follows. If independence means control, choice, and power, then fully autonomous robots may undermine the independence they claim to provide. The authors propose three modes as design metaphors — Omakase (passive following), Monitor (information-seeking dialogue), Boss (active command) — and found participants fluidly shifted between them by situation.
Autonomy is a design variable, not a goal.
Longitudinal delegation (Hata et al., 2026) is one of the few long-term evaluations of an assistive robot in the wild. Over three weeks, one participant’s delegation rate rose from 20% to 100%; another went from 0% to 54.5%. Query behaviour shifted from curiosity-driven exploration in week one to targeted, context-specific queries by week three. Delegation is not a fixed user preference. It is a dynamic negotiation shaped by context, personality, and accumulated skill in reading the robot’s state. One-shot usability studies will not catch this.
Kamikubo, Kacorri, and Asakawa’s 2024 paper — “We are at the mercy of others’ opinion” — shows navigation isn’t only about getting from A to B. Their 18 blind participants wanted a layered information architecture to support recreational window shopping: shop names pushed automatically, deeper information pulled on demand. Voice commands beat touchscreens because hands were busy with canes, harnesses, and shopping bags.
Manipulation: The Robot as Hands
Bonani et al.’s 2018 study — “What My Eyes Can’t See, A Robot Can Show Me” — compared voice-only and physically collaborative robotic assistance for blind participants assembling a Tangram puzzle. In the collaborative condition, all 12 participants completed the puzzle. In the voice-only condition, only 2 of 12 did. Eleven of twelve preferred the collaborative robot.
This is the first empirical evidence that physically collaborative robots dramatically outperform voice-only assistive technology in manipulation tasks.
Bhattacharjee et al. (2019) argued that robot-assisted feeding research has been too focused on the mechanical act of getting food to a mouth. One expert described being fed as “a degree of infantilization… it’s such a personal thing to have someone shoveling food in your mouth.” Another reminded the team: “When we feed, we talk, we laugh, we ask how they are doing.” The greatest perceived benefit of feeding robots wasn’t efficiency but independence and self-esteem. The assistive-robotics field has been too focused on engineering at the expense of human factors.
Sharma, Murthy, and Biswas (2022) used eye-gaze control of a robot arm with users experiencing severe speech and motor impairment. All 13 participants — including six quadriplegic wheelchair users — completed pick-and-place tasks safely. Eye gaze can control complex manipulation. Not just communication.
Telepresence: The Robot as Body
Friedman and Cabral (2018) gave six participants with developmental disabilities telepresence robots to tour a marine centre. Pre-tour vocabulary: “excited,” “nervous,” “tired.” Post-tour: “happy,” “impressed,” “exhilarating.” Both physical and social self-efficacy trended upward.
For people who are homebound or dependent on others for community access, telepresence can bypass transportation, ticket purchasing, crowd navigation, and physical fatigue all at once.
Socially Assistive Robots
Not every assistive robot is there to carry or guide. A distinct and very large class — socially assistive robots — provides support primarily through rapport, engagement, motivation, and conversational presence.
Olatunji et al.’s 2026 scoping review of 205 empirical studies of robots for older adults (2010–2022) found that social robots accounted for nearly half of all studies; socially-assistive hybrids another 23%; purely assistive robots only 16%. By task, over half focused on companionship and entertainment. Core activities of daily living — bathing, feeding, dressing — got only 8%. In eldercare, the field is dominated by social support, not heavy-lifting mechanical aid.
Balasuriya, Sitbon, and Mitchell (2026) compared two commercial robots with adults with intellectual disabilities. Pepper, a 120 cm humanoid, elicited immediate human-style behaviour — handshakes, introductions, affectionate goodbyes. One participant kissed Pepper’s hand. But Pepper also raised conversational expectations it couldn’t meet — one person dropped out after realising Pepper was limited to pre-set phrases. Cozmo, a small toy-like robot, produced delight because it exceeded the low expectations its form set.
The design heuristic: align appearance with capability. Don’t let the form make promises the system can’t keep.
Valencia et al.’s 2021 Socially Assistive Sidekicks paper framed robots as supporting AAC users rather than speaking for them. AAC devices generate synthesised speech, but not the nonverbal channel — gaze, pointing, gestures — that verbal speakers use to regulate conversation. Three design parameters emerged from their co-design workshop: attention (draw or maintain a partner’s attention), precision (nuance matching the user’s intention), and timing (at the exact needed moment).
Zhong et al.’s 2026 study used a Furhat robot to screen for perinatal depression, which affects up to 10% of people during pregnancy or postpartum and goes undiagnosed in 50–69% of cases. Four of five participants were open to robot-delivered screening, valuing the robot’s neutrality and non-judgement. But trust was not blanket. Critical requirements surfaced: calibrated anthropomorphism, explicit role transparency, and human complementarity with clear escalation pathways.
Piedade et al. (2024) co-designed a tabletop robotic game with 80 children (18 neurodivergent) across four Portuguese classrooms. Their central finding: the co-design process itself — not the resulting game — was the primary driver of inclusion. Haidenhofer et al.’s 2024 paper centres Chloe Haidenhofer, a woman with intellectual disability who has transitioned from research participant to co-researcher over five years. Her contribution is a graduated confidence model — talking board → robot like Pepper → peer conversation → support worker. The robot becomes a rehearsal space for human interaction.
From Autonomy to Sovereignty
Jang, Carrington, and Begel’s “From Autonomy to Sovereignty” (CHI 2026) systematically analysed 90 social-accessibility papers and argues that assistive technology has long treated “independence” as its primary goal. But disabled people’s experience is saturated with interdependence and chosen reliance.
They propose relational sovereignty — the recognised authority of disabled people to choose their relational mode and set the terms on which it is upheld. The shift is from “Can they do it?” to “Do they get to decide?”
| Independence | Interdependence | |
|---|---|---|
| Conditional | Abandonment masked as self-reliance | Coerced reliance under external gatekeeping |
| Recognised | Chosen solitude, reversible | Self-determined collective access |
Sovereignty-oriented design aims rightward — toward the “Recognised” column — regardless of relational mode. The argument lands on everything else in this article: the AI Suitcase loss of obscurity, the feeding robot concern about isolation, the Beyond Omakase modes. All instances of technology either conferring or withholding sovereignty, depending on how the terms are set.
Care and the Aging Population
Rudzicz et al. (2015) studied speech interaction between older adults with Alzheimer’s and a mobile robot. Over 40% of trouble-indicating behaviours were “lack of uptake” — participants simply ignoring the robot. Yet interactions with the robot were more likely to be successful than interactions with human conversants. Structured, task-oriented robot prompts may actually reduce certain types of confusion. But automatic speech recognition dropped to 5.8% accuracy during household tasks. Real home acoustics are far from the lab.
Frijns et al. (2024) took a different approach: instead of finished robot prototypes, they introduced technology components — computer vision, conversational AI — to 13 care workers and 12 residents. A tension worth sitting with surfaced: residents reported feeling monitored by technologies care workers viewed positively for safety.
Neurodivergent Design: Programmability Over Prescription
Kong et al. (2024) worked with eight neurodivergent participants using quadruped robot dogs. Two findings stood out. First, robot failures had outsized emotional impacts — one participant interpreted the robot falling over as personal rejection. Participants preferred robots that acknowledged failures transparently. Second, being in a caregiving role toward the robot was experienced positively — it gave participants a sense of competence. This challenges the conventional framing of assistive technology as assistance flowing in one direction.
Pulatova and Kim (2024) used programmable Toio cubes as fidget devices for 16 adults with ADHD. Traditional fidget toys offer fixed, one-size-fits-all affordances. Participants wanted state-dependent customisation — slow, rhythmic, predictable for high stress; faster and more stimulating for low focus. Programmability, not cleverer fixed behaviours, is the design lever.
Augmentation, Not Compensation
Barbareschi et al.’s 2024 Brain Body Jockey project documents a collaboration between a professional VDJ with advanced ALS and a research team. The artist uses a 16-channel EEG interface to control five robotic arms for live music performance. At a “Meet the DJ” event, over 40 fans physically interacted with him via the robotic hands — handshakes, fist-bumps, high-fives — with haptic feedback transmitting touch back to his forearm.
All 13 interviewed audience members described a feeling of fusion and connection. The artist:
“By extending my body through the power of brainwave and robotic arm technology, I can create rhythms at my own pace and encourage the audience to clap along, which gives me a greater sense of physical freedom.”
Assistive technology is too often framed as compensatory — filling functional gaps. Human augmentation can instead enable entirely new forms of creative expression. Engage with disabled artists as technological pioneers, not passive recipients of assistive solutions.
Education and Embodied Learning
RoboGraphics (Guinness et al., 2019) combines a tablet, laser-cut cardboard overlays, and small $60 Ozobot robots as moving tactile markers for dynamic data. In a seven-participant study, tactile-robot and audio-only interfaces achieved identical accuracy on bar chart comprehension — but all seven preferred the robots. The reason was gestalt understanding: touching all data points simultaneously to feel the overall shape of the data. “With the robots, you don’t have to find the data points because they are already there.”
GoDonnie (Damasio Oliveira et al., 2019) is a text-based, screen-reader-compatible programming language with spatial awareness commands — turning robot programming into a vehicle for orientation and mobility skill development.
MotionBuddy (Tanaka et al., HRI 2026) asks a different question: can a robot teach a blind learner how to move their own body? A 40 cm tabletop humanoid with tactile joint markers demonstrated Bon Odori and Karate sequences to 11 blind participants. For moves requiring simultaneous limb coordination and pose transitions, the robot produced significantly higher reproduction accuracy than narrated audio. Task-load workload was 43 for the robot versus 70 for audio.
Pragmatically: several participants framed the robot as a solution to the awkwardness of asking a human instructor for the close physical contact needed to understand posture.
Touch communicates simultaneous multi-limb relationships in a way sequential speech cannot. Neither modality alone is sufficient.
Rehabilitation and Workplace
Ng’s 2019 study of a post-stroke exoskeleton revealed a striking split. Stroke patients and caregivers rated it highly (8.0–8.9 on credibility). Biomedical researchers rated it markedly lower (4.6 credibility, 1.5 on expectancy). The patient said his hand felt “normal again.” Clinical effectiveness and user satisfaction can diverge. Patient voices need to be in the evaluation loop from the start.
Heinz-Jakobs et al.’s 2025 AT@Work paper names the “disability divide” — the sociotechnical disparity in access to digital technologies. Collaborative robots are well accepted by workers with disabilities, but many workplace ATs are still designed without meaningful input from the people who use them.
Urban Infrastructure
Froehlich et al. (2024) maps where robotics meets the built environment. Project Sidewalk has collected over 500,000 crowdsourced accessibility labels. Over 600,000 people with disabilities in the US alone cannot leave their homes due to transportation barriers — yet autonomous vehicle development has largely ignored accessibility in vehicle design, pickup logistics, and interaction modalities. Emerging transportation technologies must be designed with disability from the outset, not retrofitted later.
Cross-Cutting Themes
Across all 36 papers, several patterns recur.
Embodiment matters. Physical collaboration outperforms verbal instruction. Haptic feedback deepens connection. Being in the space — even via a robot body — reshapes participation.
Programmability beats prescription. Fixed affordances fail neurodivergent users, fixed feeding patterns ignore individual pacing, fixed rehabilitation movements ignore individual capacity. The robots that work adapt.
Autonomy is a design variable, not a goal. Full autonomy can paradoxically strip agency. The design work is in legibility and fluid mode transitions.
Sovereignty, not just autonomy. What matters is not whether the robot makes users independent, but whether they get to set the terms on which independence or interdependence operates.
Expectation must match capability. Anthropomorphic form makes promises the system must keep. Calibrate anthropomorphism downward to what the system can actually deliver.
Failure communication is design. How a robot signals that it’s struggling is as much a design concern as how it behaves when it’s working.
Longitudinal changes everything. One-shot studies systematically understate how delegation, trust, and query patterns shift over weeks of real use.
Visibility is a design choice. Making someone blend in is not always a kindness. Assistive technology alters how disabled people are perceived by strangers.
Participatory design is not optional. Technically excellent robots fail without the people who will use them shaping the design from the start.
Augmentation is a legitimate frame. The compensatory framing — robot as crutch — is too narrow. Robots can extend bodies, enable new creative expression, and support identity rather than merely compensate for impairment.
Where This Points
Robotics is one of the few areas in accessibility technology where the hard problems of physical presence can actually be addressed. Navigating unfamiliar spaces. Manipulating objects. Teaching a dance pose. Occupying a room. Offering a calm non-judgemental presence through a difficult conversation.
None of the papers above suggest a near-term future of humanoid household assistants. They suggest something more tractable and more human: specific, narrow, well-designed robots that do one or two things extraordinarily well, built with the people who will use them, honest about their failures, and understood as partners rather than replacements.
The AI Suitcase research programme — tracked across eight years from CaBot to daily museum operation to longitudinal delegation studies — is the clearest existence proof in the field. It shows what a sustained interdisciplinary programme of work with blind users can achieve when it survives contact with real public spaces.
Two takeaways.
If you work in accessibility and you have written off robotics as a distant concern, the research is reason to look again.
If you work in robotics and you have not built with disabled users, the evidence is clear: you will build the wrong thing.
Sources: 36 peer-reviewed papers from ASSETS, CHI, HRI, ACM Transactions on Human-Robot Interaction, TEI, W4A, TACCESS, SIGACCESS Accessibility and Computing, and PETRA, published between 2007 and 2026.