RIC: Can Robots Handle Infants?

Roughly 500,000 US infants are admitted to a neonatal intensive care unit each year, into care that is among the most physically and cognitively demanding in medicine, and among the least automated. The World Health Organization projects a shortage of 11 million healthcare workers by 2030, and in the NICU specifically, burnout reaches 37%.

Our clinical partners have already started asking when they can have a humanoid robot of their own. So we asked the prerequisite question: can current robots physically handle an infant safely enough to help?

This project is in active development. The paper is targeted at ICRA 2027; the results and figures below are work in progress.

Why nobody had answered it

Existing automation in infant care monitors babies or passively soothes them. Nothing touches them. The obstacle isn’t the motion, it’s the tolerances. A neonate withstands a fraction of the mechanical load an adult can, and the head must stay within a narrow angular range throughout a lift or the airway is compromised. Those limits are clinically specified and unforgiving, and they are the reason infant handling has stayed manual.

We picked the two interventions that matter most:

Bimanual pickup. Infants in the NICU are handled for roughly 2.5 hours every day, and a pickup bookends nearly every one of those episodes: weighing, feeding, imaging, line management, and repositioning all begin and end with the infant being lifted.

CPAP nasal-mask repositioning. CPAP is the first-line noninvasive therapy for preterm infants in respiratory distress, and nasal masks need frequent nurse-initiated repositioning (our clinical partners reported up to 50 times a day) to preserve the seal. Each adjustment is a force balance: too little and the seal leaks; too much and contact pressure causes nasal skin breakdown, which affects a majority of infants on long-duration support.

The two fail in completely different ways: inertial and postural on one side, sustained contact pressure on the other. That is exactly why a safety methodology for infant handling has to generalize beyond a single task.

Top: head–torso pitch beyond the clinical limit, compromising the airway. Bottom: a lift within limits, the hand supporting both neck and back.

Turning clinical judgment into numbers

I worked with four neonatal clinicians (a NICU nurse, a NICU physician, and two neonatal respiratory therapists) plus the literature they pointed us to, converting “safe handling” into quantities you can measure continuously during a trial:

• Pickup: cervical pitch (the head–torso angle that governs airway patency) and head acceleration.
• CPAP: interface pressure at the nasal bridge and philtrum, geometric exclusion zones around the eyes and mouth, and a hard time budget on PEEP interruption, since even brief losses of positive pressure risk lung de-recruitment.

Measuring what a robot actually does

We benchmark three ways of performing each task, direct human handling, teleoperation, and an autonomous learned policy (ACT), on a Unitree G1 humanoid, using an instrumented infant manikin tracked by OptiTrack, and a Fisher & Paykel nasal mask I rebuilt with embedded force sensors to read interface pressure at the contact points clinicians actually worry about. Every trial is verified against the thresholds above, so “safe” is a measurement rather than an impression.

The pipeline: OptiTrack tracks the manikin and computes the safety quantities; teleoperated demonstrations become training data; the learned policy runs on the humanoid and is scored against the same thresholds.

The safety map on the infant's face: red exclusion zones over the eyes and mouth, yellow pressure-sensitive areas, and the green nasal target the mask must reach.

What went wrong, and what it taught us

The first hands we tried failed in a way that turned out to be the most informative result of the project. A person lifting an infant works their fingers underneath and wiggles them to break contact friction with the bedding before committing to the lift, which is what lets the hand reach a supporting position under the head and back before any weight transfers. Our candidate robot hands could not do this: they lacked the degrees of freedom, and their quasi-backdrivable spring mechanisms caught on the bedding, making frictional sticking a dominant failure mode.

We replaced them with smooth, low-friction plastic hands with no added degrees of freedom, trading dexterity for reliable entry. That forced a scoop strategy rather than pre-shaping around the infant, and it explains why cervical pitch, not head acceleration, is the binding safety constraint: a hand that cannot reach a supporting position before load transfer lifts with less of the head and back supported. The acceleration data confirms the failures are not caused by jerky motion. They are caused by the configuration the robot can reach before the lift begins, which points the next iteration at wrist force-torque sensing and compliant actuation rather than at smoother trajectories.

Project lead and first author; targeted at ICRA 2027.