Signals InboxJuly 16, 2026Humanoid Robotics
Humanoids: what healthcare tasks can they do today?
Humanoids can already talk with patients, run structured assessments, guide rehabilitation exercises and handle simple bedside deliveries. Their ability drops fast when the work involves lifting a person, using medical tools precisely or deciding what treatment a patient needs.
We track humanoid robotics daily. Want the market signals in your inbox?
Send me the signals →Humanoids can already perform useful healthcare tasks today, but mostly when the interaction is scripted, supervised and easy to stop. Patient conversation, cognitive exercises, hospital navigation, simple deliveries and clinician-controlled examinations are currently the strongest use cases.
The best evidence is not coming from robotic surgery. It comes from quieter workflows: Robin operating across 30 US healthcare facilities, Pepper administering a standardized assessment to 100 patients and Florence completing medication, delivery and vital-sign workflows in a ward involving 67 patients.
There is a hard autonomy divide. Humanoids can navigate, ask fixed questions and coach exercises with limited assistance. Once sustained physical contact, tool use or clinical interpretation begins, teleoperation and direct human supervision take over.
Repetition helps robots; variation breaks them. Ventilation timing, questionnaires and rehabilitation instructions fit machines well because the protocol is already known. Patient transfers, suturing and emergency procedures expose weaknesses in force control, dexterity, tactile sensing and failure recovery.
The near-term product is therefore closer to a robotic assistant than a robotic nurse. Clinicians choose the protocol and manage exceptions, while the humanoid handles the repetitive, physical or conversational part.
Interested in humanoid robotics?We can send you all the signals
Send me the signals → Delivered straight to your inboxQ1What healthcare tasks can humanoids actually do with patients today?
Humanoids currently work best with patients when the interaction can be scripted, supervised and stopped without causing physical harm. Social engagement, structured assessments and bedside deliveries account for 13 of the top 15 tasks in our ranking.
The scale difference is striking. Robin has reached 30 healthcare facilities, Pepper has completed a randomized assessment study involving 100 patients, and Florence was evaluated with 67 hospital patients. None of those programs asked the robot to lift a person, insert a medical device or decide on treatment.
Physical care is progressing much more slowly. The strongest patient-transfer evidence consists of nine staged demonstrations over three days, while the main direct-intervention study used simulators and remote operators. Healthcare humanoids have achieved meaningful patient exposure, but mostly in tasks where clinicians can easily supervise the interaction and reverse an error.
Humanoid healthcare tasks ranked by current capability
| # | Healthcare task | Category | Capability score | Evidence stage | Control mode | What has actually been proven |
|---|---|---|---|---|---|---|
| 1 | Holding personalized conversations with children during hospital stays | Patient engagement | 82 | Operational | Hybrid | Robin is currently used in 30 US healthcare facilities. It remembers patient preferences, plays games and provides emotional support, but only around 30% of its behavior is autonomous; remote operators still handle much of the interaction. |
| 2 | Playing personalized games and music with hospitalized children | Patient engagement | 80 | Operational | Hybrid | Robin conducts repeated bedside activities based on children’s names and preferences. The multi-facility footprint is unusually strong for a humanoid healthcare program, although published clinical outcome data remains limited. |
| 3 | Leading cognitive and social activities for dementia residents | Elderly care | 76 | Patient pilot | Autonomous or supervised | Pepper has led structured cognitive training in dementia care, while Abi is now entering a year-long UC Davis study involving up to 25 residents. The field is moving from short acceptability sessions toward longer observation of whether engagement survives the novelty period. |
| 4 | Administering a standardized cognitive screening questionnaire | Clinical assessment | 74 | Patient pilot | Autonomous interview | Pepper administered the MMSE in a randomized study involving 100 patients with mild cognitive impairment and was compared directly with neuropsychologist-led assessment. The robot conducted the structured test; it did not make a broader diagnosis. |
| 5 | Delivering preloaded medication to an identified hospital patient | Nursing support | 72 | Patient pilot | Autonomous navigation | Florence located patients and presented medication from a secure compartment during a real medical-ward study involving 67 patients. Staff retained responsibility for selecting, preparing and loading the medication. |
| 6 | Delivering drinks and small personal items to hospital bedsides | Nursing support | 71 | Patient pilot | Autonomous navigation | Florence completed bedside deliveries in the same ward study. This workflow carries less clinical risk than medication delivery and mostly combines navigation, patient identification and object presentation. |
| 7 | Locating and identifying a known patient inside a hospital ward | Nursing support | 69 | Patient pilot | Autonomous | Florence navigated the ward and identified enrolled patients before starting assigned workflows. Its task was bounded to known participants and a mapped environment. |
| 8 | Presenting devices for bedside vital-sign measurements | Physical examination | 68 | Patient pilot | Patient-assisted | Florence presented measurement equipment for blood pressure, temperature and oxygen saturation. Patients or staff still had to use or position parts of the equipment, so the robot facilitated collection rather than independently performing a full examination. |
| 9 | Guiding patients through structured rehabilitation exercises | Rehabilitation | 65 | Patient pilot | Autonomous coaching | Socially assistive robots have delivered exercise instructions and feedback across multi-session rehabilitation programs. A long-term post-stroke study reported better rehabilitation outcomes with robotic support, moving the evidence beyond one-session acceptability testing. |
| 10 | Monitoring whether rehabilitation movements match prescribed exercises | Rehabilitation | 62 | Repeated | Autonomous | Experimental systems can compare patient movements with expected trajectories and flag incorrect execution. Their role is currently closer to exercise supervision than clinical evaluation or treatment planning. |
| 11 | Giving corrective verbal feedback during rehabilitation exercises | Rehabilitation | 61 | Repeated | Autonomous | Rehabilitation robots can repeat instructions and corrections without fatigue. The strongest value appears during high-repetition sessions, while therapists continue to define the program and interpret progress. |
| 12 | Greeting visitors and explaining hospital services | Navigation and administration | 59 | Patient pilot | Autonomous | A review of 26 hospital studies found social robots performing greeting, education, companionship, goods transport and staff-training functions. Greeting is among the easiest to repeat because it involves limited physical risk. |
| 13 | Collecting structured patient information through spoken questions | Clinical administration | 58 | Patient pilot | Autonomous interview | Pepper’s 100-patient cognitive-assessment study shows that humanoids can administer fixed clinical questionnaires consistently. Their current strength is following a validated script, not deciding which questions a complex patient needs. |
| 14 | Explaining an upcoming medical procedure to a child | Patient engagement | 57 | Operational | Hybrid | Robin has been used to model procedures such as IV placement and help children understand what will happen. Human clinical staff decide what should be explained and remain present for the procedure. |
| 15 | Leading breathing and relaxation exercises during patient distress | Patient engagement | 56 | Operational | Hybrid | Robin has guided dementia residents and distressed patients through breathing exercises and personalized calming activities. The reported value is emotional de-escalation rather than treatment of the underlying condition. |
| 16 | Transferring a person between a wheelchair and a bed | Patient mobility | 53 | Repeated | Hybrid | RHP Friends demonstrated patient-transfer workflows three times daily over three days at IREX. The nine repeated demonstrations are stronger than a single video, but they involved a staged environment and combined autonomous actions with teleoperation. |
| 17 | Positioning a digital stethoscope on known chest landmarks | Physical examination | 49 | Repeated | Teleoperated | A Unitree G1 reproduced the positioning movements for cardiac auscultation. Robot vibration degraded the acoustic signal, showing that reaching the correct location does not guarantee clinically usable data. |
| 18 | Applying controlled pressure with a stethoscope or examination tool | Physical examination | 48 | Repeated | Teleoperated | The G1 used impedance control to regulate tool contact, but sensor sensitivity remained insufficient for dependable clinical interpretation. |
| 19 | Holding an oxygen mask securely against a patient simulator | Emergency care | 47 | Repeated | Teleoperated | The G1 coordinated both arms to maintain a bag-valve-mask seal on a simulator. The operator selected the placement and corrected the robot remotely. |
| 20 | Squeezing a ventilation bag at consistent programmed intervals | Emergency care | 46 | Repeated | Programmed | Ventilation timing was more consistent than the human comparator in simulator testing. This is one of the few medical tasks where repetition suits the robot particularly well, although mask placement and patient assessment still require human control. |
| 21 | Palpating an abdomen through a predefined examination sequence | Physical examination | 43 | Demonstrated | Teleoperated | The G1 reproduced the four Leopold maneuvers on a maternal simulator. Its hands could follow the sequence but lacked the tactile sensitivity needed to interpret what a clinician would feel. |
| 22 | Holding an ultrasound probe against a tissue phantom | Medical imaging | 42 | Repeated | Teleoperated | The robot maintained probe contact using force-aware control while an operator searched for a simulated vessel. This proves remote positioning more clearly than autonomous imaging. |
| 23 | Aligning an injection needle with an ultrasound image plane | Medical intervention | 40 | Repeated | Teleoperated | A remote operator coordinated the probe and needle during 20 simulated injection trials. The humanoid supplied the physical interface, while the operator interpreted the ultrasound image. |
| 24 | Inserting a needle into an ultrasound-visible simulated vessel | Medical intervention | 39 | Repeated | Teleoperated | The study reported 45% direct hits and 70% success after adjustments across 20 trials, compared with a cited 90% benchmark for experienced clinicians. The result shows feasibility but also a meaningful accuracy gap. |
| 25 | Inserting an endotracheal tube into an airway simulator | Emergency care | 36 | Repeated | Teleoperated with assistance | The robot completed the broad movement sequence, but a human had to apply extra force with the laryngoscope. The humanoid could reach the airway without independently overcoming the procedure’s physical resistance. |
| 26 | Cutting a simulated neck incision for emergency airway access | Emergency care | 31 | Repeated | Teleoperated | Only 30% of ten simulated tracheostomy cuts were fully successful; another 40% were partial and 30% failed. This remains an experimental capability with an error rate incompatible with clinical use. |
| 27 | Inserting a tracheostomy tube into a simulated airway opening | Emergency care | 30 | Repeated | Teleoperated | The G1 aligned and partly inserted the tube but struggled with the required force and dexterity. The experiment exposed a mechanical limitation rather than a lack of medical planning. |
| 28 | Driving a curved needle through a surgical training pad | Medical intervention | 29 | Repeated | Teleoperated | Across 16 attempted suture throws, only 43.8% of complete attempts succeeded. Needle insertion caused most failures, placing basic humanoid suturing well below clinical reliability. |
| 29 | Releasing and regrasping a needle after tissue penetration | Medical intervention | 28 | Repeated | Teleoperated | The robot completed individual clamp and regrasping motions more reliably than the full suturing sequence. Chaining the movements together created most of the difficulty. |
| 30 | Pouring a measured drink for a person requiring assistance | Daily care | 26 | Demonstrated | Programmed | Nursing-care prototypes have demonstrated quantitative pouring in controlled environments. Evidence involving varied containers, patient movement and real swallowing risk remains limited. |
| 31 | Helping a patient eat a complete meal safely | Daily care | 20 | Demonstrated | Supervised | Feeding prototypes can move food toward a known target, but complete meal assistance adds food variation, swallowing risk, patient communication and continuous correction. |
| 32 | Repositioning an immobile patient safely inside a hospital bed | Patient mobility | 16 | Claimed | Not proved | No humanoid has publicly demonstrated reliable autonomous repositioning of a real immobile patient while managing limbs, tubes, pressure points and discomfort. |
| 33 | Bathing a patient while preserving safety and dignity | Daily care | 12 | Claimed | Not proved | Bathing appears regularly in nursing-robot roadmaps, but current humanoids have not proved the full workflow with real patients. Water, privacy, balance and skin contact compound the difficulty. |
| 34 | Recognizing patient deterioration and independently starting treatment | Clinical decision-making | 6 | Claimed | Not proved | Current humanoids can collect scripted information or act as remote physical surrogates. None has proved that it can combine symptoms, vital signs and patient history to initiate treatment independently. |
| 35 | Independently managing all physical and clinical care for one patient | Complete patient care | 1 | Claimed | Not proved | No humanoid currently combines medication, mobility, hygiene, examination, emotional support and emergency judgment without continuous human responsibility. |
Interested in humanoid robotics?We can send you all the signals
Send me the signals → Delivered straight to your inboxFigure AI launches Index with 16 million videos already uploaded
Chinese robot cut its 100m record from 9.39 to 8.86 seconds
XPeng’s robot unit raises $900M before selling a single robot
Unitree’s robot IPO jumps 460% in Shanghai
Unitree CEO sees robots hitting "a ChatGPT moment" by 2028
Unitree’s Superman robot hits 12.66 m/s, faster than Bolt’s peak
FCC blocks Unitree’s next robots, opening America’s humanoid race
San Francisco residents can now hire a humanoid cleaner for $30
Google DeepMind gives robots a Gemini brain that acts
Musk: it will be "America’s fastest industrial scale-up since WWII"
BotQ just built Figure’s 1,000th F.03 humanoid robot\
Tesla is building its first Optimus lines
Q2Are healthcare humanoids autonomous or still teleoperated?
Healthcare humanoids are currently autonomous in conversation, navigation and fixed questionnaires. Once sustained physical contact begins, human control takes over.
The split appears almost perfectly in the ranking. The top 15 tasks can generally run through autonomous navigation, scripted interaction or a hybrid system. Every task from stethoscope placement through suturing relies on teleoperation, programming or direct human assistance.
Robin also shows why the word “autonomous” needs care. It works across 30 facilities, yet around 70% of its behavior still involves remote control.
Recent technical progress has improved the quality of that remote control. Whole-body systems can now map human motion to a Unitree G1 with low latency, and medical researchers have combined teleoperation with impedance control to handle tools more safely. These advances make humanoids better physical surrogates, but they do not transfer clinical judgment to the machine.
The market is dividing into two products: relatively autonomous social robots and clinician-controlled humanoid bodies. A robot that can independently combine both roles has not emerged yet.
Q3Which healthcare tasks are closest to real deployment?
Structured cognitive assessment, bedside delivery and rehabilitation coaching are the most credible near-term healthcare tasks now. Each already has evidence beyond a laboratory demonstration, and none requires a robot to make an irreversible medical decision.
Pepper’s 100-patient cognitive study is particularly instructive. The robot could administer a standardized test because the questions, scoring framework and escalation path were already defined. Florence followed the same design logic in a ward: staff prepared the medication or equipment, while the robot handled navigation, identification and presentation.
Rehabilitation adds a different advantage. Repetition is expensive for human therapists but cheap for a robot. Long-term research now suggests that robotic coaching can improve post-stroke rehabilitation rather than merely entertain patients during a single session.
The therapist still chooses the exercises and evaluates progress. The robot gets the repetitive part.
These three use cases share the same architecture: clinicians define the protocol, the robot repeats it, and humans retain authority over interpretation and exceptions. That pattern looks much more deployable than a robotic nurse performing broad bedside care.
We track humanoid robotics daily. Want the market signals in your inbox?
Send me the signals →Q4Which healthcare jobs are humanoids still nowhere near doing?
Humanoids remain far from replacing nurses, doctors or caregivers because complete healthcare jobs combine routine execution with constant exception handling. The robot performs much better when the correct action has already been selected.
The seven-procedure Unitree study makes the gap unusually measurable. Ultrasound-guided vessel insertion reached 70% after adjustments, simulated tracheostomy cuts succeeded fully in 30% of trials, and complete suture throws succeeded 43.8% of the time. Auscultation produced another type of failure: the arm reached the correct chest positions, but robot vibration polluted the signal.
Those failures came from four different bottlenecks: accuracy, force, dexterity and sensor quality. Better language models alone will not solve them.
Patient care also adds shifting body positions, discomfort, tubes, fluids and sudden deterioration. None of those appeared together in these controlled tests.
Recent healthcare research has proved that humanoids can carry medical tools and reproduce a clinician’s movements. It has not shown that they can understand a changing patient, recognize when the planned action has become unsafe and choose a better one.
We track humanoid robotics daily. Want the market signals in your inbox?
Send me the signals →The question “What healthcare tasks can humanoids actually do today?” is difficult to answer from isolated demonstrations. A robot may perform one impressive movement without completing the broader task, working autonomously or operating safely around real patients. We therefore broke the question into specific healthcare tasks and assessed the evidence behind each one.
We ranked each complete healthcare task rather than the underlying ability to speak, navigate or move an arm. The score combines task completion, autonomy, repetition, exposure to real patients and operational scale.
For each task, we reviewed recent deployments, patient studies, repeated experiments and controlled demonstrations. We prioritized evidence showing what the robot actually completed, under which conditions, with how much human control and with what measurable result.
We then compared the strongest evidence across different robots and studies. This allowed us to identify broader patterns without treating one impressive demonstration as representative of the whole field.
We also separated physical execution from clinical authority. A humanoid transporting medication, asking a standardized questionnaire or positioning a medical instrument may complete an important part of the workflow while clinicians still select the action, interpret the result and manage exceptions.
Claimed means the complete task remains an ambition. Demonstrated means it was performed in a controlled environment. Repeated means it was reproduced across several trials. Patient pilot means real patients or care residents were involved. Operational means the task is used repeatedly inside healthcare facilities.
No humanoid healthcare task currently meets our threshold for scaled deployment. The ranking reflects the strength of the evidence available today, not the task’s technical sophistication or long-term potential.
A relatively simple task can rank highly when it has been repeated with real patients across healthcare facilities. A technically advanced procedure can rank lower when it remains teleoperated, simulator-based or unreliable.
Key sources used for this analysis include: AP News on Robin’s use across 30 US healthcare facilities, its approximately 30% autonomy and its patient-engagement activities, research on humanoid performance across seven medical procedures, including auscultation, ventilation, ultrasound-guided injection, intubation, tracheostomy and suturing, research on RHP Friends’ repeated patient-transfer demonstration, research on personalized robotic monitoring and corrective feedback for post-stroke rehabilitation, research on personalized socially assistive robot interactions for stroke rehabilitation, a review of socially assistive robot deployments across healthcare environments, research mapping autonomous and teleoperated robot use in clinical care and healthcare logistics, research on Pepper’s participatory deployment with children in a real care and education environment, and research on the design and clinical evaluation of a socially assistive robot for older adults with low vision.
Building or investing in humanoid robotics?We can send you all the signals
Send me the signals → Delivered straight to your inbox