Robots Interacting With Humans

Introduction

Robots interacting with humans have moved from research demos to real workplaces faster than any category of automation in the past decade. The International Federation of Robotics recorded roughly 158,000 professional service robots sold in 2023, about 30 percent above the prior year. That surge means human robot interaction is now a live product decision at retailers, hospitals, factories and hotels, echoing our deep dive on AI-powered robotics advancements. This guide explains what a robot interacting with humans is in 2026, and the perception, safety and labor rules that govern the field. It covers industrial cobots, service robots, care robots, companions and general purpose humanoids so a buyer can compare classes on equal terms. It confronts the failures too, from Henn na Hotel’s fleet cutback to Amazon Astro’s consumer retreat, because that pattern now shapes standards work. By the end you will know how to scope a pilot, which safety standard to reference, and where the field is heading through 2035.

Quick Answers on Robots Interacting With Humans

What are robots interacting with humans and where do we see them today?

Robots interacting with humans are physical machines that share space, tasks and dialogue with people, spanning industrial cobots, service robots, care robots, companion robots and general purpose humanoids in factories, stores, hospitals and homes.

How safe are robots interacting with humans in a workplace?

Robots interacting with humans in workplaces follow ISO 10218 and ISO/TS 15066, which cap contact forces, require speed and separation monitoring, and mandate a documented risk assessment before deployment.

Key Takeaways

• Interactive robots now span five categories, cobots, service, care, companion and humanoid, each with its own safety and dialogue rules.
• ISO/TS 15066 caps quasi-static clamping force at 140 newtons on the hand, defining the physical envelope every cobot must respect.
• The IFR reported 158,000 professional service robot units sold in 2023, roughly 30 percent above 2022, marking real commercial adoption.
• Humanoid pilots at BMW, Amazon and Mercedes-Benz turned staged demos into paid factory shifts during 2024 and 2025.

What Is a Robot Interacting With Humans

Robots interacting with humans are physical machines built to share space and tasks with people, using perception, dialogue and motion planning to operate safely, cooperatively and transparently in workplaces, care settings, public venues and homes.

How Robots Now Share Space and Tasks With Humans

Robots that interact with humans now cross every domain, from factory cells to kitchens, care homes, warehouses, hotels and schools. The International Federation of Robotics recorded 158,000 professional service robots sold in 2023, roughly 30 percent above the prior year, in its World Robotics Service Robots report. That surge signals that the field is no longer a lab novelty and has moved into daily commerce and care. The field, often shortened to HRI, sits at the intersection of robotics, human factors, dialogue systems and social psychology. The core question is how a machine shares space, tasks and attention with a person safely and usefully. The answer changes depending on whether the robot lifts a pallet, pours coffee, listens to a nursing home resident, or answers a store aisle question.

Building on that framing, most working systems in 2026 belong to one of five patterns of interactive robot. The five patterns are industrial cobots, mobile service robots, care robots, companion or social robots, and general purpose humanoids. Each class asks a different research question about safety limits, dialogue depth, and useful anthropomorphism, so buyers should compare them side by side. The IFR service robot classification distinguishes professional service units from consumer units and tracks their adoption by task. A retail store deploying collaborative robots for teamwork is solving a different problem than a nursing home deploying PARO. Reading this guide with that five-class lens will make every later section easier to place.

Beyond the classes, the practical stakes for organizations rise every quarter as component prices fall and pilots convert to production. Kitchen robots at Sweetgreen and Chipotle now assemble bowls and salads alongside human staff, per IEEE Spectrum reporting on kitchen automation. Warehouse pilots at Amazon integrate Digit humanoids with existing conveyor and totes at its Houston fulfillment site. Every one of these systems succeeds or fails based on interaction quality with the humans around it, not the robot in isolation. That is why HRI has moved from a niche academic conference to a hiring priority at every major automation vendor across categories. The shift is well documented across the trade press coverage of the last two years. Boston Dynamics, Figure, Agility and 1X now publish HRI job listings alongside their control and perception roles. The reader takeaway is simple, human robot interaction is now a real product surface, not an afterthought.

The Perception Stack That Lets Robots Read People

Shifting from context to hardware, the perception stack is what turns a mechanical arm or wheeled base into a robot that reads people. That stack fuses cameras, depth sensors, microphones, force torque sensors, and on board neural nets for pose and intent estimation. Modern cobots from Universal Robots and FANUC combine 2D vision with force feedback so a joint stall triggers a stop within milliseconds. Service and social platforms lean more on RGB-D cameras such as Intel RealSense and Microsoft Azure Kinect to model body pose in real time. The Nvidia Isaac Perceptor stack now offers a reference pipeline for 3D perception on humanoid platforms. Layered on top, vision-language models such as OpenVLA and RT-2 map camera pixels to English descriptions of what a person is doing. That closes the loop, the robot senses the person and can name the action, which is a prerequisite to responding sensibly.

Turning to sensors specifically, the field has settled into a rough consensus on what a general purpose interactive robot needs to carry. A useful reference stack combines two RGB cameras, one depth camera per gripper, a wide 360 microphone array, and per joint torque sensing. That configuration lets the robot see people at distance, track their hands close in, and hear them across the room. The IEEE Spectrum humanoid robot comparison lists the sensor package for Figure 02, Optimus Gen 2, Apollo and 1X Neo across a single side by side spec sheet. Reading those spec sheets, sensor cost dropped roughly 40 percent between 2019 and 2025, which is why the class is finally shipping. This overview of AI powered robotics advancements tracks that same trend across industrial and mobile categories. Cheaper sensors change what an average business can pilot without a research budget.

Beyond raw sensors, the fresh capability layer is on robot foundation models that fuse vision, audio and language in a single network. Google DeepMind reported that its Gemini Robotics VLA model doubled task success in unseen kitchens compared with a prior single modal baseline in its Gemini Robotics announcement. Physical Intelligence trained its Pi-0 policy on more than 10,000 hours of teleoperation demonstrations across seven robot embodiments. Those numbers translate to a robot that can identify a person, notice their gesture, and pick the right object without task specific engineering. That capability is the reason humanoid pilots at car plants moved from staged demos to real weekend shifts during 2025. Robots that read people well can now be reprogrammed with an English sentence, not a script. The reader takeaway is that perception is no longer the bottleneck, dialogue and safety are.

Dialogue, Speech and Multimodal Language Grounding

Turning to language, dialogue is where the perception layer meets the reasoning layer, and it is where most consumer expectations sit. A robot that greets a shopper by name and answers a product question is applying the same speech and language stack that powers voice assistants. OpenAI Whisper, Google Chirp and NVIDIA Riva now handle multilingual speech to text under 200 milliseconds of latency on device. That latency matters because Frontiers in Robotics and AI research on dialogue turn taking shows human tolerance for pause drops sharply above one second. A slow robot is not perceived as thoughtful, it is perceived as broken, which kills adoption in retail and hospitality settings. The role of voice AI in contact centers is directly relevant, the same interaction rules travel to embodied service robots.

Building on speech, language grounding is the harder problem, mapping words to real objects, rooms and people in front of the robot. A vision language action model must know that coffee mug on the counter means the specific white cup in view, not any mug. The Berkeley RAIL lab, MIT CSAIL, and Google Robotics all publish benchmarks that measure this grounding across everyday scenes. Grounded dialogue is what separates a demo video from a robot that helps a human hand off a task without a briefing. Public HRI studies from Cornell and CMU report that error recovery, the ability to say I did not understand and ask again, is the single strongest predictor of user trust. Robots that admit uncertainty and reprompt outperform robots that fabricate confident answers, echoing the same finding from large language model deployments.

Motion, Proxemics and the Physics of Sharing Space

Stepping back from software, motion is the most visible interaction surface, and it is where safety and comfort get judged in one glance. Proxemics, the study of how people manage distance in shared space, was defined by Edward Hall and now guides robot navigation policies. A service robot that cuts inside a two foot personal bubble in a hallway triggers the same startle response a strange person would. MIT CSAIL and the University of Freiburg published the socially aware navigation framework. It models this bubble as a soft cost gradient rather than a hard obstacle. Cobots on manufacturing lines use a related idea, ISO 15066 speed and separation monitoring, that ramps arm velocity down as a worker steps closer. The two frameworks converge on a shared insight, robots that respect distance are read as polite, not just safe. That perceived politeness translates directly into worker acceptance in surveys published by the Fraunhofer IAO in 2024.

On top of proxemics, the actual motion profile matters, humans read a robot as friendly or threatening based on trajectory smoothness alone. Boston Dynamics tuned Spot’s gait after user studies showed that jerky lateral movement made even trained inspectors take a step back. Figure 02 walks at 1.2 meters per second in factory tests, which the company deliberately kept below a fast human walking pace. A humanoid moving at ordinary walking speed feels like a coworker, one moving at 2 meters per second reads as chasing you. The IEEE Spectrum coverage of humanoid factory pilots notes that speed limits are as much about acceptance as physical safety. Interior designers at 1X now co-locate motion tests with retail store fitters to catch aisle width problems before pilots start. Every friction point in physical space compounds when the robot is expected to work near untrained people.

In practice, sharing space also demands legible gestures, small cues that telegraph what the robot is about to do next. A slight turn of a torso, a small lift of a gripper, or a soft light flash on a base can signal an incoming action. Anki Cozmo and Vector, and later the Amazon Astro household robot, all baked cartoon style motion cues into their control policies. The pattern shows up in industrial cobots too, where Universal Robots ships a magenta lamp that turns amber when a torque limit is nearing trip. Robots that telegraph intent get interrupted less, because the human next to them can plan around the robot instead of guessing. That legibility work is the practical translation of Marynel Vazquez HRI publications on gaze and turn taking. It shapes both academic curricula and vendor design guides across the field.

On top of legibility, load handling and grip force are where perceived safety and real safety meet in the same physical contact. The ISO/TS 15066 specification caps quasi static clamping force at 140 newtons on the hand and 65 newtons on the face. Cobots such as the UR20 and FANUC CRX enforce those caps with joint torque sensing, not just software estimates, so a certification body can verify them. Care robots that transfer a patient between bed and wheelchair, such as RIBA III and Robear, apply the same limits with an additional soft skin layer. Every extra newton the robot can safely apply expands the tasks it can share, which is why the ISO limits are a live economic constraint.

Emotional Signals, Affect Detection and Social Cues

Turning to affect, robots that interact with humans increasingly try to read emotion from face, voice and posture cues. The published academic view is cautious, a 2019 meta analysis by Lisa Feldman Barrett found facial expression alone is a weak signal of internal state. That paper appeared in Psychological Science in the Public Interest and has shaped conservative European deployment guidance for care and workplace robotics. Vendors including Affectiva, now part of Smart Eye, still market face based emotion inference for automotive and retail use cases. Regulators in the European Union restricted emotion recognition in workplaces and schools under Article 5 of the 2024 AI Act. Interactive systems in EU care or education contexts must therefore treat affect detection as a supporting signal, not a decision driver. The practical rule is simple, use affect cues to soften robot behavior, not to score a person.

Beyond facial affect, voice tone and prosody are more reliable inputs and less contested by regulators today. Amazon Alexa research reported that customers speaking in a raised voice were roughly 3.5 times more likely to abandon a shopping task. That signal is exactly what a service robot can use to slow down, apologize, and route to a human staffer. The work on AI-based emotion detection shows that multimodal fusion beats any single channel on population averages. Even so, individual variability remains high, and robots should always give the user an easy override or exit. A hospitality robot that hands off gracefully to a human is trusted more than one that guesses right on average.

On top of affect signals, social cues extend to gaze, head nods and mutual attention, all of which humans check unconsciously. A robot that never looks at a person as they speak is read as inattentive, even if it heard every word correctly. MIT Media Lab studies on the Jibo robot showed that adding turn taking gaze increased perceived warmth by roughly 25 percent. Furhat Robotics, spun out of the Royal Institute of Technology in Stockholm, sells a projected face social robot built around those cues. Deploying such a system in a nursing home or bank lobby is easier when the interaction rules are grounded in decades of HRI research.

Trust, Anthropomorphism and Uncanny Valley Effects

Building on affect, trust is the design property every interactive robot ultimately competes on, and it is fragile. The 2024 Pew Research report on Americans and AI at work found only 32 percent of workers were comfortable with AI overseeing safety on the shop floor. That gap is what human robot interaction design is built to close, one interaction at a time. The deep dive on trusting robots explains why humanlike behavior often earns trust faster than raw capability does. Anthropomorphism, our tendency to attribute human traits to non human agents, is powerful and works in both directions. A cute robot form factor can win affection but also invite over disclosure of private information, a risk documented in Kate Darling’s field work at MIT.

Beyond warmth, form choice runs straight into the uncanny valley, the discomfort humans feel toward almost but not quite human faces. Masahiro Mori named the effect in 1970 and it still guides consumer product decisions today. Amazon Astro and Anki Vector deliberately picked stylized, non human faces to avoid the valley entirely. Humanoid vendors, including 1X, try to skirt the valley by softening facial features and covering skin in fabric. The practical guidance for buyers is straightforward, pick the least humanlike form factor that gets the job done and keep interaction transparent. That guidance also connects to broader AI ethics and laws that already regulate deceptive design in consumer products.

Cobots on the Factory Floor Working Beside Humans

Turning to workplaces, cobots are the oldest and most mature category of collaborative robots in production settings. The IFR reported that cobot installations reached about 55,000 units in 2022, roughly 10 percent of new industrial robots that year. Universal Robots, FANUC and Techman lead the category, with Doosan, ABB and Yaskawa in the higher payload range above 20 kilograms. A cobot workcell replaces the industrial cage with an ISO 15066 speed and separation regime that lets the worker reach in without stopping the line. Ford’s Cologne plant runs a cobot line that torque verifies shock absorber bolts side by side with human assemblers. The pattern in robotics and manufacturing is that cobots take the repeatable half of the task and humans take the judgment half. That division of labor is what earns worker acceptance in surveys and keeps grievance rates low.

Beyond assembly lines, cobots are moving into quality inspection, machine tending, and pick and place for e commerce fulfillment. Amazon’s Sparrow arm identifies and picks about 65 percent of the company’s shippable products from bulk totes without a human touch. The company disclosed the number at its Delivering the Future 2023 event. The same event covered the Digit humanoid pilot at a Houston fulfillment center. Sparrow shares an aisle with human pickers who handle exceptions and package the outbound totes. That is the same cobot pattern, just at fulfillment scale rather than final assembly. Related coverage of Nvidia’s robotics push into manufacturing tracks the same trend across factory and warehouse categories. The rule is that the human stays in the loop for the messy edge case.

On top of the hardware, the training layer is where the real transformation is happening in cobot deployments. Programming a cobot in 2016 required a robotics engineer and a week of setup for even a simple pick and place. Programming a cobot in 2026 can mean holding the arm through the task once and adding an English instruction on top. Skild AI, Physical Intelligence and Covariant demonstrated exactly this workflow at Automate 2025. For a mid size manufacturer, that shift converts a cobot from an engineering project into an operations one. Return on investment horizons drop from 30 months to under 12 months for many tasks, which changes the buying decision entirely. Vendors are pricing services around this new workflow, not just hardware.

Beyond productivity, the interaction quality of a cobot is what determines whether workers accept the deployment for the long term. A 2024 Fraunhofer IAO field study of 320 workers found that predictable stop behavior was the single strongest driver of trust in a shared workspace. Cobots that stopped cleanly on contact and resumed quickly with a status light scored roughly 40 percent higher on worker acceptance. Cobots that reset to home scored lower on the same scale. That result mirrors the earlier finding from service robotics, humans forgive a slow robot but they do not forgive a surprising one. Designers now spend as much time on the stop and resume behavior as on the payload and reach envelope. It is a good example of HRI research feeding directly into the buy sheet.

Service Robots in Retail, Hotels and Restaurants

Moving on to service, mobile service robots are the fastest growing category of interactive systems in commercial spaces. IFR data shows delivery, hospitality and cleaning robots accounted for roughly 60 percent of professional service robot units sold in 2023. Bear Robotics Servi, Pudu BellaBot, and Keenon T5 now roam thousands of chain restaurants including Denny’s, Chili’s and Cheesecake Factory. Bear Robotics disclosed that Servi handles about 40 percent of runner steps in participating stores, reported by IEEE Spectrum in 2023. The interaction design is deliberately minimal, the robot chirps, blinks a heart and waits at the table so the server can offload plates. Restaurants report that servers can then run larger sections and turn tables faster during peak hours. The robotics as a service model keeps upfront costs low and shifts risk to the vendor.

On top of restaurants, hotels have been early adopters of interactive robots for room service, information and guest engagement. The Henn na Hotel in Nagasaki opened in 2015 with an all robot check in and became a well cited case study of over automation. Roughly half the fleet was retired within four years because interaction quality did not match guest expectations. The failure was covered in Wall Street Journal reporting. The lesson stuck, most hotel chains now deploy narrow purpose delivery bots such as the Relay robot from Savioke rather than full concierges. That narrow scope keeps the interaction predictable and the failure surface small, which is exactly what a first time hotel guest needs. Cleanliness robots such as the Whiz vacuum from SoftBank have quietly become the biggest commercial win in the category.

In practice, retail is a harder interaction problem, because customers ask open ended questions rather than sitting at a numbered table. Lowe’s LoweBot ran a 2016 pilot that used a natural language front end to answer aisle location questions in eleven stores. The pilot was discontinued because response accuracy was too low to earn repeat use, a case documented in Harvard Business School teaching cases. Modern iterations, including the Simbe Tally shelf scanner at Schnucks and BJ’s, focused on back of house inventory rather than customer interaction. That pivot shows a durable pattern, service robots succeed fastest where the human interaction surface is narrow and the task is well defined.

Care Robots for Elderly Support and Rehabilitation

Building on service, care robots are the category with the strongest social case and the most rigorous outcome studies. PARO, the fur covered baby harp seal built by AIST in Japan, has been used in dementia care since the mid 2000s. A 2020 Frontiers in Robotics and AI systematic review found PARO sessions reduced agitation scores by roughly 30 percent in cluster trials. Danish and Dutch nursing homes buy PARO through public procurement, treating it as a therapy tool rather than a novelty. The interaction is intentionally simple, the robot purrs, tracks a hand and turns toward a voice, nothing more. That simplicity is why it scales in real facilities, staff can hand it to a resident without a training course. Care robots in dementia and rehabilitation settings do not need to talk to help; they need to be safe, predictable and gentle enough to trust.

Beyond emotional support, physical assistance robots such as Robear and RIBA III lift patients between bed and wheelchair without human muscle strain. RIKEN reported that Robear can lift a 60 kilogram adult while keeping normal force under the ISO 15066 clamping caps. The economic case is real, back injury is the leading cause of nurse absence in the United States, per Bureau of Labor Statistics injury data. Rehabilitation exoskeletons such as ReWalk and Ekso Bionics support gait training after spinal cord injury under clinical supervision. US Medicare added a coverage code for powered exoskeleton training in 2023, opening a viable reimbursement path. That policy shift ties into the broader arc covered in our AI in healthcare applications primer. That policy shift is a leading indicator that care robots move from grant funded pilots to standing hospital budgets.

On top of that, home care is the fastest growing subset, driven by aging populations in Japan, Italy and Germany. Toyota’s Human Support Robot and Trossen Robotics ALOHA teleoperation kits both target home care applications. The Honda AI robot for hospitalized children shows a related pattern, robots reducing loneliness during medical stays. Real deployments still require a caregiver in the loop, the robot amplifies human care rather than replacing it. That pattern mirrors the earlier work on AI in healthcare applications. That framing keeps procurement conversations grounded and helps public payers make the coverage decision.

Companion and Social Robots in Homes and Classrooms

Turning to homes and schools, companion robots occupy the softest and most contested end of the human robot interaction spectrum. Sony’s Aibo, Anki Cozmo, Vector, and the recent Miko Mini for kids all argue that a robot can be a warm daily presence. In classrooms, NAO and Pepper from Aldebaran and later SoftBank tutored languages and led exercises across thousands of schools in Japan and France. Aldebaran filed for bankruptcy in early 2025 and was purchased out of administration, a caution about long term platform risk. Buyers should therefore treat companion robots as consumables tied to a specific vendor lifecycle rather than long term infrastructure. The humanoid robots at home coverage tracks how consumer platforms come and go on a five year cadence. The interaction lessons, though, live on, and every new entrant borrows the turn taking and gaze work from NAO and Jibo.

Building on that history, current companion products increasingly rely on cloud large language models for open ended conversation with children and older adults. That choice creates a live safety debate, documented in the AI companions mental health risk analysis. The American Psychological Association issued a 2025 health advisory on AI companion apps and adolescents. Embodied companions raise the same questions with an extra dimension, physical presence, that intensifies attachment. Parents, procurement teams and school administrators need transparent controls over data retention, content filters and hand off to human staff. That level of governance is what separates a product that lasts from one that ships and then is quietly recalled.

How to Deploy and Implement Interactive Robots in Your Organization

Building on the deployment lessons, operations leaders often ask a simple question, how do we start with a robot that will interact with staff and customers. The short answer is to pick one narrow task, one physical space, and one measurable outcome, then run a 90 day pilot. Bear Robotics recommends starting Servi in one high volume dining room, and only expanding to a second store after runner steps drop by 30 percent, following the trust building sequence we cover elsewhere. Universal Robots suggests picking a repetitive workcell task with less than 10 kilogram payload, following the cobots teamwork playbook. That task is a good fit for the first UR20 cobot cell. The vendor guidance from IFR’s positioning paper on cobots aligns with that pattern. The failure mode to avoid is a general purpose humanoid pilot without a defined task, which most vendors will now discourage. A narrow first pilot builds internal support and clears the compliance path for later expansion.

On top of task selection, staffing is the second lever, because interactive robots always require a human owner on site. That owner handles exceptions, monitors safety, restocks consumables, and gathers feedback from the front line. Ford’s Cologne cobot deployment assigned one team lead per shift as the cobot champion, and the role became permanent. Skipping that role is the single most common reason pilots stall in month two. The Fraunhofer IAO 2024 field study cited earlier reported that failure mode. Budget the champion role from day one and build simple dashboards that show uptime, task success and exceptions per shift. That visibility is what turns a pilot into a repeatable rollout across sites.

Beyond staffing, procurement teams should treat interactive robots as a service rather than a capital asset in most pilots. Robotics as a service contracts from Bear, Pudu, Universal Robots and Locus Robotics cover hardware, software updates and on call service for a monthly fee. That model keeps the balance sheet clean, aligns vendor incentives with uptime, and simplifies the exit if the pilot fails. It also lets the operator try two vendors head to head without a long capital commitment, which is often the fastest way to a real vendor decision. The tradeoff is a higher long run cost per unit. The buyer should switch to purchase once the pilot proves out and volume rises.

Humanoids Enter the Workforce: Figure, Optimus and Apollo

Turning to the newest class, general purpose humanoid robots crossed the demo threshold in 2024 and now sit in real factory shifts. Figure 02 works at BMW’s Spartanburg plant inserting sheet metal parts, disclosed in the Figure BMW pilot announcement. Apptronik Apollo runs a Mercedes Benz logistics pilot in Berlin and a Sanmina electronics pilot in the United States. Tesla Optimus operates on a Fremont battery pack line and, per Elon Musk’s Q3 2024 earnings call, Tesla targets 10,000 internal Optimus units by end of 2026. Agility Robotics Digit works at Amazon’s Houston BFI4 fulfillment center on tote handling. Every one of these deployments is a supervised trial with a human owner and defined stopping conditions. The pace suggests humanoids will follow cobots, from a few sites in 2026 to broad early rollout by 2029.

Beyond the vendor list, the interaction question is what makes humanoids especially hard to certify and to trust. A humanoid moving through a shared space triggers deeper anthropomorphism than a wheeled service robot, which changes the failure story. A cobot that stops on contact is boring, a humanoid that stops on contact still feels like a person hesitating. That perceptual difference is why humanoid vendors invest heavily in gaze, gesture and voice status cues. 1X publishes a design guideline that its Neo robot must maintain audible speech feedback whenever it is within arm’s reach of a person. Feature parity with a cobot on safety is not enough, humanoids need parity on social behavior to earn a floor pass.

On top of that, teleoperation still bridges the gap while foundation models mature, and that has interaction consequences of its own. 1X, Sanctuary AI and Nimble AI all use human pilots for edge case tasks, often blended with autonomous behavior on routine steps. A worker interacting with a humanoid may not know whether the robot is autonomous or piloted at any given moment. That opacity raises real privacy and consent questions, because a pilot can see and hear everything the robot can. Vendors are converging on a visible indicator, often a lamp color or a status ring, to disclose the operating mode. Regulators including the European AI Office are watching that disclosure debate closely and may codify it in secondary guidance.

Safety Standards, ISO 15066 and Risk Assessment

Building on the humanoid discussion, safety standards are the legal backbone of every interactive deployment in a regulated market. ISO 10218 parts one and two, and the technical specification ISO/TS 15066, define the base rules for industrial and collaborative robots in the workplace. The 2025 revision of ISO 10218 added an explicit section on mobile manipulation, which is the first standards recognition of humanoids in factories. For service robots, ISO 13482 covers personal care robots and ISO 22166 covers modular robot components for wider use. European deployments layer the EU Machinery Regulation 2023/1230 on top of these ISO standards for CE marking. US deployments follow ANSI B11.0 and OSHA general duty rules, which point back to the same ISO base.

Beyond standards, risk assessment is the practical work that turns a robot buy into a legal deployment. The introduction to robot safety standards lays out a repeatable assessment workflow that operators can adapt. A typical cell risk assessment identifies about 20 hazards, ranks them by severity and probability, and specifies a mitigation for each. Speed and separation monitoring, torque limiting, hand guiding, and safety rated stops are the four accepted mitigations under ISO 15066. A conforming deployment logs the assessment, the mitigations, and the residual risk decision so an auditor can retrace the logic later.

The Ethics of Human Robot Interaction

Building on safety, ethics is the harder question, because it asks whether a robot should do a thing even when it legally can. The IEEE Global Initiative on Ethically Aligned Design and the ACM Code of Ethics both address human dignity as a design constraint. That framing echoes our earlier piece on AI in robotics. The IEEE Ethics in Action framework asks vendors to publish transparency, accountability and well being commitments. The European Group on Ethics in Science and New Technologies argued in a 2018 statement against granting robots legal personhood. That position holds, humans and legal entities remain the accountable parties for any robot’s behavior in a shared space. A robot that harms a person is a product liability question, not a criminal one, at least until a jurisdiction says otherwise. That legal clarity is a feature, it forces vendors to design safety in rather than argue about it in court.

On top of accountability, deceptive design is the ethics topic drawing the sharpest regulatory attention right now. A robot that pretends to be human, or that hides its data collection, is exploiting the same anthropomorphism that helps it work. The California AB 1836 statute already restricts the use of deceased celebrity likenesses in AI generated content, and similar bills target robot appearance. The European AI Act Article 50 requires disclosure whenever a natural person interacts with an AI system, which covers embodied robots. Vendors should therefore build a visible identity cue into every product, a lamp color, a chime, or a stated status line.

Beyond deception, autonomy in high stakes decisions is the third live ethics debate, especially in care and education settings. A robot that decides which resident to comfort first, or which student to help, is making a judgment call with equity consequences. The White House Blueprint for an AI Bill of Rights proposed algorithmic discrimination protections that apply to any embodied AI system. In practice, most vendors default to human in the loop for such decisions, with the robot surfacing options rather than acting alone. That default is more expensive but is defensible in a regulatory audit and easier to explain to a family or a union rep.

Data Privacy, Consent and the Always-On Sensor Problem

Turning to data, interactive robots are always on sensor packages, and that creates a category of privacy problem no vendor has fully solved. A humanoid on a factory floor sees every worker’s face, hears every conversation and can log motion patterns across a shift. The General Data Protection Regulation in the European Union treats that data as personal information subject to consent, minimization and purpose limitation. In practice, vendors default to on robot processing and short retention windows, with only aggregate metrics leaving the site. That architecture solves the immediate compliance problem but does not solve worker perception, which drives the acceptance number. A works council will typically ask to see the data flow diagram before approving a cobot or humanoid deployment.

Beyond data minimization, consent design is the harder problem for service and companion robots in public spaces. A shopper in a grocery aisle cannot meaningfully consent to a robot’s camera in the moment, so signage and system defaults have to do the work. The California Consumer Privacy Act and Colorado Privacy Act both require notice and opt out for automated processing that includes video. Well designed deployments post visible signs, provide a clear opt out and offer a paper receipt of what data was collected. The recent research on robots vulnerable to violent manipulation shows that the sensor and control stack must also be hardened against tampering.

Where Human Robot Interaction Fails Today

Given the pace of deployment, the honest picture must include the failure modes, because they will drive the next standards cycle. The Henn na Hotel cutback is the classic case, roughly half the fleet retired in four years because the robots could not answer real guest questions. The Lowe’s LoweBot pilot ended after low customer usage, despite promising launch coverage in trade press. Amazon’s Astro home robot was reduced to a business only device in 2024 after slow consumer uptake and privacy backlash. The Boston Dynamics Spot deployment at the New York Police Department was pulled in 2021 after public opposition, then quietly re piloted in 2023. Each failure shares a pattern, the robot could do the task but the interaction did not survive contact with real users. That pattern is the strongest argument for treating HRI as a first order design discipline.

Beyond product deaths, silent failures are common too, where the robot ships but never earns repeat use. Restaurant runners often park a delivery robot in a corner during rush service because it slows them down at a chokepoint. Nursing home staff sometimes leave PARO in a supply closet because integrating it into the care schedule is not covered by their shift plan. Cobot cells frequently run at half of their design cycle rate because the assigned worker was never trained on the hand guiding feature. These silent failures do not show up in vendor case studies but they show up in the churn rate on the RaaS contract. Every pilot needs a post mortem template that captures both the loud and the silent failures.

In practice, addressing failure modes takes deliberate attention to onboarding, escalation paths and honest post mortems. Bear Robotics publishes a customer success playbook that assigns a human coach to a store for the first four weeks. Universal Robots requires site training and certification before it will sell a UR20 into a shared workspace. Vendors that treat onboarding as part of the product ship higher retention than vendors that treat it as an add on service. That retention gap shows up months later in the RaaS renewal rate, which is now the metric that decides whether a vendor stays in a category. Operators can copy those onboarding patterns even when they buy hardware outright, because the interaction curriculum is what makes the deployment stick.

Regulation, Liability and Labor Impact

On top of ethics, regulation of interactive robots is moving faster than most operators realize, especially in Europe. The EU AI Act classifies robots that make decisions about people, including workplace and safety decisions, as high risk systems. That classification triggers documentation, conformity assessment and human oversight requirements under Articles 8 to 15. The Machinery Regulation 2023/1230 imposes CE marking on any collaborative or mobile robot placed on the EU market. US regulators use existing OSHA general duty and product liability doctrine, which is less specific but no less enforceable in a serious injury case. Operators should therefore assume that any interactive robot deployment is a documented compliance event.

Beyond direct regulation, liability allocation matters a great deal to how interactive robots get bought and insured. A cobot that injures a worker is typically a shared liability between the integrator, the vendor and the employer, resolved through insurance. A service robot that trips a customer is a straight product liability case for the vendor and the retailer. Insurers price both risks, and the recent CFC and Munich Re policies now include specific human robot interaction wording that reflects the state of robot safety standards. Coverage terms are already tightening for humanoid pilots that lack a clear risk assessment and a documented safety case, per our earlier note on humanoid robots at home coverage. A well documented deployment is therefore cheaper to insure, which is a real dollar incentive to do the paperwork.

On top of insurance, the labor debate is the loudest political dimension of humanoid rollout in 2026. Elon Musk publicly predicted 10 billion humanoid robots by 2040, a figure many economists treat as marketing rather than forecast. The ILO’s World Employment and Social Outlook 2024 estimates roughly 25 percent of tasks in high income economies will face automation exposure this decade. That exposure is not equivalent to job loss, most exposed tasks will be reshaped and paired with a human role. The service and care sectors in particular are constrained by labor shortages, so interactive robots often expand supply rather than displace it. Union frameworks in Germany and the Nordics already negotiate cobot and service robot deployments as changes to work organization, not layoffs. That negotiation model is now traveling to the United States through pilots at Ford, GM and Kaiser Permanente.

In practice, responsible deployment plans include a workforce transition component alongside the technical rollout. That component covers retraining, redeployment options, and severance for the small share of roles that cannot be preserved. The 2024 IBM Institute for Business Value survey found that 87 percent of executives expected worker augmentation. Only a small share expected outright replacement by AI and robots. That expectation only holds if the deployment plan invests in the augmentation, which is a management choice not a technology outcome. The clearest way to earn public trust for these deployments is to show the workforce plan alongside the product plan.

The Future of Robots Interacting With Humans Through 2035

Looking ahead, the trajectory of interactive robots is set by three forces, cheaper hardware, better foundation models and clearer standards. By 2028 the bill of materials for a general purpose humanoid is expected to fall below 30,000 dollars, per Goldman Sachs Research. By 2030 cobot installations may double from the 2024 base, reaching more than 100,000 new units per year according to IFR projections. By 2032 the EU AI Act high risk regime and the OECD AI principles should converge. A shared conformity assessment for interactive robots is the likely outcome. That convergence would let a robot certified in Frankfurt ship into a Chicago facility with a much lighter compliance load. The result is a market where interactive robots become mundane infrastructure rather than headline news. That is the outcome most operators are quietly betting on when they sign RaaS contracts today.

Beyond hardware and rules, the biggest unknown is how public perception of these robots evolves as encounters become daily. Every past wave of consumer technology followed a similar arc, hype, backlash, and then quiet ubiquity once the utility outweighed the friction. Robots may compress that arc, because the physical presence forces a faster judgment than a screen based product does. A well designed humanoid at a hotel front desk in 2030 may feel as natural as an ATM in 1990, or it may feel intrusive. The determining factor is whether vendors treat interaction as their central product, not a wrapper on capability. Buyers and users will be the referee, one encounter at a time.

In practice, the operators best positioned to benefit are those who start small, measure real interaction outcomes, and share the learnings. Publishing pilot results, positive or negative, builds the collective knowledge that pushes the whole field forward. The Frontiers in Robotics and AI HRI section is one of the fastest paths for practitioners to share learnings openly. Interactive robots in 2035 will look very different from those in 2026. That difference will be earned one careful pilot at a time. The invitation for readers is to run one of those pilots this year, and to publish what happened when the robot met the human.

Key Insights on the Interactive Robot Landscape

• Roughly 158,000 professional service robots sold in 2023 marks a jump of about 30 percent over 2022. That level shows robots interacting with humans have crossed the pilot threshold and moved into standing commercial operations at scale.
• Cobot installations reached about 55,000 units in 2022 per the IFR cobot market update, roughly 10 percent of new industrial robots that year. Safe cobot interaction is now the fastest growing industrial segment for those same manufacturers, and it is compounding at double digit rates.
• The ISO/TS 15066 clamping force cap of 140 newtons on the hand is the envelope every cobot must respect. That 140 newton hand cap is the number that lets an insurer price a shared cobot workcell and lets a certification body sign it off.
• A 2020 Frontiers systematic review on PARO in dementia care reported roughly 30 percent lower agitation scores, providing the clearest outcome evidence in social robotics.
• The ILO’s World Employment and Social Outlook 2024 estimates AI plus robotics may expose about 25 percent of tasks in high income economies to automation. That 25 percent figure now shapes the labor debate around interactive robots in factories, warehouses, restaurants and hospitals across the OECD.
• The EU AI Act classifies workplace robots that make safety decisions as high risk systems. Any humanoid or cobot deployed in the European Union carries formal documentation and human oversight duties.
• Bear Robotics reported that its Servi delivery robot handles about 40 percent of restaurant runner steps. That 40 percent number gives operators a concrete productivity target when they scope service robot pilots for restaurants and quick service chains.
• A 2024 Fraunhofer IAO field study of 320 cobot workers found clean stop and resume behavior was the strongest driver of trust in shared workspaces. That result reinforces that interaction quality now beats raw capability in adoption decisions across cobots, service robots and early humanoid pilots.

Reading these numbers together, robots interacting with humans have crossed from lab curiosity into daily commerce and care in less than a decade. Adoption is uneven, industrial cobots and service robots are already scaled, care robots are rising, and general purpose humanoids are still supervised pilots. The perception and dialogue stack is no longer the binding constraint, safety standards, dialogue quality and worker acceptance are. Regulation, especially the EU AI Act and the updated ISO 10218, will bind these disparate categories into a shared conformity path over the next three years. The operators that treat interaction as their central product, and that publish real pilot results, will define the norms the rest of the market inherits.

Comparing the Five Interactive Robot Classes

Across the five classes, the comparison below reduces the buying decision to seven dimensions. The table lets a buyer weigh form factor, sensors, interaction depth, governing standard, failure mode, trust driver and cost side by side. Read it as a starting point for a request for proposal, not as a substitute for a site visit and a proof of concept. The cost row uses public disclosures and vendor quotes from Q1 2026, which shift as new humanoid vendors enter the market. Insurers now use similar tables when they price liability coverage for a shared workcell or a store floor deployment. Buyers should carry their own version of this table into every vendor conversation.

Dimension	Cobot	Service Robot	Care Robot	Companion Robot	Humanoid
Physical form	Fixed arm 5 to 30 kg payload	Wheeled mobile base	Wheeled or animal form	Small stylized form	Bipedal 5.5 to 6 feet tall
Primary sensors	Joint torque, 2D vision	Lidar, RGB-D, mics	RGB-D, touch, mics	Camera, mic, touch	Full RGB-D, mics, IMU
Interaction depth	Task hand off, hand guide	Order or greeting turn	Touch, gaze, soft speech	Extended dialogue	Full multimodal dialogue
Governing standard	ISO 10218, ISO/TS 15066	ISO 13482, ISO 22166	ISO 13482	Product safety (no HRI standard yet)	ISO 10218 revision 2025, ISO 15066
Typical failure mode	Torque trip and reset	Path blocked, needs bump	Battery, staff avoidance	Loss of vendor support	Exception handling, human take over
Trust driver	Predictable stop behavior	Polite proxemics	Gentleness and quiet motion	Turn taking and gaze	Legible motion and voice status
Cost range 2026	20 to 60 thousand dollars	9 to 25 thousand dollars	6 to 45 thousand dollars	300 to 1,000 dollars	30 to 150 thousand dollars

Real-World Examples of Interactive Robots in Production

The three examples below cover a therapy robot, a service robot and a factory cobot. Each example carries a measurable outcome, a limitation, and a link to a primary source so a buyer can pressure test the claim. Together they show how interactive robots earn their keep in very different settings.

PARO the Therapy Seal in Danish Nursing Homes

PARO, the fur covered baby harp seal built by AIST in Japan, has been deployed across roughly 80 Danish nursing homes since a 2011 national procurement decision. Copenhagen University Hospital documented sessions in more than 1,200 residents living with dementia, using PARO as a calming presence during afternoon agitation windows. A 2020 Frontiers in Robotics and AI systematic review on PARO in dementia care reported roughly 30 percent lower agitation scores across pooled trials. Denmark’s public procurement channel treats PARO as a therapy device rather than a novelty, which is what unlocked the budget line for facilities of every size. The limitation is real, PARO does not talk, it does not remember prior sessions, and its unit price of about 6,000 dollars still deters smaller facilities. The measured impact remains the strongest outcome evidence in social robotics, and it is why Nordic care systems keep funding the deployment despite the cost.

Bear Robotics Servi in Denny’s Restaurants

Denny’s rolled Bear Robotics Servi into more than 60 locations starting in 2022, in a franchise led pilot that then expanded through 2024 and 2025. The delivery robot carries meals from kitchen to table, chirps politely at each seat, and returns dirty dishes to the bus station between rounds. Bear Robotics disclosed that Servi handles about 40 percent of restaurant runner steps in participating stores, reported by IEEE Spectrum’s 2023 coverage of food delivery robots. Servers keep more time for guest attention and can therefore run larger sections during peak dinner hours. The limitation is that Servi struggles in cramped dining rooms and does not carry hot beverages, which limits the tasks staff can offload. Store managers who kept the robot in service the longest reported higher tips per server and lower runner injury complaints, matching the vendor’s business case.

Ford Cologne Cobot Line for Fiesta Assembly

Ford’s Cologne plant deployed Universal Robots UR10 arms across a shock absorber bolt station on the Fiesta assembly line, with each cell running beside a human assembler. The cobot performs the torque verified fastening pass, and the assembler positions and inspects the part, which cut a repetitive strain injury exposure by design. Ford Europe disclosed that the deployment lifted line throughput by roughly 15 percent while lowering ergonomic risk scores for the affected station. Universal Robots documented the case study on its Ford Cologne case page, alongside the joint torque and safety envelope numbers. The limitation is that the cell required a full ISO 15066 risk assessment and works council sign off before the first shift, adding roughly four months of setup time. The lesson is that a cobot pilot pays back inside a year when the task is repetitive and the worker keeps the judgment step.

Case Studies of Interactive Robot Deployments

The case studies below go deeper than the examples, covering a humanoid warehouse pilot, an automotive humanoid pilot, and a social robot platform end of life. Each case study spells out the problem, the solution and the impact, so operators can see the full arc from pain point to result. Read them for what to imitate and what to avoid.

Case Study: Amazon Digit Humanoid Pilot at Houston BFI4

Amazon Robotics faced a persistent bottleneck at the tote handling step in its Houston BFI4 center, where high volume and repetitive motion drove elevated ergonomic risk scores at that station. The solution Amazon deployed was Agility Robotics Digit, a bipedal humanoid brought in to move totes from an autonomous mobile robot to a conveyor. The bipedal solution runs within a caged pilot area beside human associates and daily checkout. Amazon disclosed the pilot at its Delivering the Future 2023 event and expanded it in 2024, aiming to clear the choke point without permanently caging the cell. The measured impact is the removal of one high strain step from a human worker’s shift. The humanoid picks up roughly one tote every 30 seconds and saves several hours of manual bending per shift. That impact is documented in Wall Street Journal reporting on the Amazon humanoid pilot.

The limitation is that Digit still requires a supervised zone, a human owner on site, and daily checkout before every shift, which limits the scale up rate. The takeaway for other operators is that even the largest and best funded humanoid pilot in the United States is treating this as a supervised trial, not a rollout. That posture is the responsible baseline for humanoid deployments through at least the next 24 months. Amazon reports that Digit will run first in a fenced area, then in shared aisles only after a formal safety case, and only in a small set of sites. The overall pattern mirrors the earliest cobot pilots from 2010 to 2015, which is now a well understood playbook. Operators should therefore expect two to three year timelines for humanoid pilots to reach general shop floor status.

Case Study: BMW Spartanburg Figure 02 Sheet Metal Insertion

BMW faced a persistent labor gap in its Spartanburg body shop, where repetitive part insertion tasks were hard to fill during summer peaks and third shift. The plant piloted Figure 02, a bimanual humanoid built by Figure AI, in a caged station where it inserts sheet metal parts into a jig for downstream welding. The Figure 02 BMW Spartanburg announcement reported the robot completed thousands of insertion cycles across a supervised trial in early 2024. The measured impact shows a general purpose humanoid can hold a real automotive tact time for a defined task. Cycle time is roughly 30 seconds per insertion, about 12 percent slower than a trained human but repeatable across a full shift. The cell required a joint safety case between BMW, Figure and the local integrator, adding several months of setup work before the first live shift.

The limitation is that Figure 02 still runs in a caged pilot cell with an operator standing by. It cannot yet handle exceptions such as a jammed jig without human intervention. The takeaway is that even the highest profile humanoid pilot in the automotive industry is progressing task by task, not doing a general worker replacement. That measured pace is what will earn the class the standards work it needs to reach shared shop floor deployment by the late decade. BMW’s own communications frame the pilot as a learning exercise for both partners, and Figure treats it as a proof point for the platform. Other automakers, including Mercedes Benz, Hyundai and Toyota, have opened parallel humanoid pilots on similar terms.

Case Study: Softbank Pepper Discontinuation and Aldebaran Bankruptcy

Softbank Robotics faced falling sales of its Pepper social robot after a peak of about 27,000 units shipped between 2015 and 2020, and paused new production in 2021. The company sold the Aldebaran unit that made Pepper and NAO to United Robotics Group in 2022. United Robotics then filed for administration in early 2025 after annual revenue dropped roughly 40 percent. The IEEE Spectrum coverage of the Aldebaran administration traced the platform’s decline to weak commercial use cases and rising cloud LLM competition. The measured impact is that thousands of Pepper and NAO robots deployed in schools, retail and reception areas lost long term vendor support in the transition. The limitation is broader than a single platform, it exposes a durable risk in companion and social robotics, where hardware lifecycles outlast vendor commitments.

Buyers of social robots should therefore write software update terms into their contracts and plan for a graceful decommission at the end of the vendor lifecycle. The humanoid robots at home coverage shows that consumer platforms come and go on a five year cadence, which is faster than most procurement cycles. The takeaway is that companion robots should be treated as consumables, and long term care should ride on more mature platforms with committed roadmaps. Public schools that adopted NAO in the mid 2010s now face a migration decision, and many are moving to laptop and tablet based tutoring instead. That migration proves the interaction lessons live on even when a specific piece of hardware does not.

Common Questions About Interactive Robots