The Physical AI deployment paradox
Why the ‘clinical metre’ of hospital logistics is the next frontier for Physical AI
Render from Great Ormond Street Children’s Hospital digital twin
Let’s start with a blunt reality check on demographics. Increased life expectancy is colliding with a sharp global decline in birth rates. The UK’s National Health Service (NHS) already faces 100,000 open positions, and the WHO projects an 11 million global shortage of healthcare workers by 2030. The numbers simply don't add up, forcing a rethink of long-held assumptions about care.
At Apian, we see care providers turning to automation to bridge this gap. The goal isn't to replace humans; it is to take the ‘robotics’ out of human workflows, allowing staff to focus on patient care.
Our drone delivery networks achieve a 15x reduction in journey times between Guy’s and St Thomas’ hospitals. But flying across London highlighted a deeper operational challenge: today’s logistics still rely heavily on ‘human middleware’ to load, fetch, and carry samples. This burden is staggering when you consider the NHS alone processes more than one billion pathology tests each year. If highly trained clinicians are still racking up tens of thousands of steps each day carrying blood and medications, we won’t have addressed the problem.
We have solved the clinical mile, now our journey continues to the clinical metre.
The same logistics challenge that led us to drones now brings us to robotics. To solve this, we need autonomous systems capable of navigating frenetic hospital environments safely alongside patients and clinicians.
Scaling beyond the operating theatre
When people think of healthcare robotics, they usually picture surgical arms in a controlled, sterilised operating theatre. Apian is thinking differently: tackling the miles of corridors head on, with hospital-scale Physical AI.
A hospital is a sprawling, multi-layered ecosystem with heavy traffic, shifting obstacles, and complex navigation requirements. Unlike warehouses, hospitals are designed for humans rather than robots. Corridors become crowded, equipment moves constantly, workflows change throughout the day, and safety requirements are uncompromising.
While traditional robotics takes us some of the way there, bridging the gap between today’s AI models and the physical world is what NVIDIA calls the $50 trillion opportunity. In healthcare, the convergence can’t come soon enough.
But hospitals aren't places to ‘move fast and break things’. Unlike web applications, failures cannot simply be patched after deployment. In a hospital, mistakes happen around patients.
This is the Physical AI deployment paradox: autonomous systems need real-world experience to become safe, but they must be safe before they can gain real-world experience.
To break this paradox, we need digital twins.
Here’s a look at how we did it.
The multi-layer control stack: from SLAM to VLAs
We employ a rigorous, multi-layer control stack. At the base, we use mature, deterministic methods like SLAM (Simultaneous Localisation and Mapping). These systems safely and reliably handle navigation along hospital logistics routes.
However, the ever-changing nature of a modern hospital requires greater capability to handle complex environments over time. To navigate previously unmapped environments, interact naturally in a lift, or seamlessly adapt to unexpected situations like patient emergencies, we are bridging the gap with the latest AI: Vision-Language-Action (VLA) models.
Apian robot simulated inside Great Ormond Street Children’s Hospital digital twin
The physics of the twin
The challenge is that VLA models require vast amounts of experiential data to operate safely - data that cannot be ethically gathered through trial-and-error in a live clinical setting. For us, a digital twin is much more than just a staging environment; it is a physically accurate simulation where these advanced navigation models can be tested safely thousands of times.
Apian has joined the NVIDIA Inception programme to accelerate our digital twinning work in healthcare. We are modelling world-class institutions like Great Ormond Street Hospital and Guy’s and St Thomas’ NHS Foundation Trust, creating simulated environments to rigorously train and evaluate future robotic systems before they ever hit the real world.
With Project Rheo, NVIDIA’s hospital automation blueprint within Isaac for Healthcare, we built this infrastructure, taking high-fidelity multi-camera photogrammetry surveys, fused with LiDAR to give rich, highly detailed data, accurate to millimetre scale. This yields dense point clouds of more than 82 million points in total, which are then stitched into a unified coordinate frame, representing a skeleton of the entire space.
At the cutting edge: balancing precision, photorealism, and scale
Processing this sheer volume of raw data in extreme parallelisation requires enterprise-grade GPUs. Simulation forces a choice between fidelity and scalability, and our goal is to achieve the physical and structural accuracy required for training our logistics robots at scale. As Physical AI and VLAs rapidly evolve, their tolerance for simulation artifacts decreases; they demand increasingly higher fidelity to successfully transfer learning from simulation to reality: the so-called ‘Sim-to-Real’ gap.
To stay ahead of this shifting bar, our pipeline leverages the NVIDIA Omniverse. Rather than relying on heavy, manually asset-mapped meshes, we train 3D Gaussian Splatting (3DGS) models through NVIDIA’s Neural Reconstruction (NuRec) pipeline. This captures complex physical spaces as millions of view-dependent ‘splats’. By running the rendering pipeline via a custom renderer in a CUDA container on NVIDIA A100 GPUs, we ensure extremely high-throughput data processing. This lightweight but highly accurate reconstructed geometry feeds directly into NVIDIA Isaac for Healthcare, allowing robotic agents to master navigating intricate hospital environments entirely in simulation.
Sovereign infrastructure for whole-hospital Physical AI
By mastering this simulation environment, we unlock ever more capable real-world deployments. We believe that this framework will drive Physical AI adoption across healthcare systems at rapid scale.
Our initial use cases focus on moving pathology samples, blood products, and urgent medications. Successfully addressing the pathology use case alone across 50% of English NHS hospitals could free enough staff time to unlock more than 2 million additional inpatient interactions annually.
A CI pipeline for improved care
A static model is only the beginning. Using Isaac for Healthcare, we are bringing these twins to life with interactive elements like functioning doors, virtual lifts, and dynamic obstacles such as unpredictable pedestrian traffic. We are programming high-stakes edge cases: dynamic rerouting around out-of-service lifts, navigating spillages, and reacting to local emergency protocols and Standard Operating Procedures (SOPs).
The Physical AI data flywheel
The multi-layered approach enables a virtuous cycle, or a ‘data flywheel’. The traditional navigation stack operates the robot today, allowing us to map facilities and compile detailed digital twins.
Crucially, Apian’s robots detect people’s presence using local AI models and automatically mask captured pixels from their cameras, before data is ever transferred or stored. This anonymisation at the edge allows safe recording of the hospital’s layout, while preserving the privacy of patients, staff and visitors. Because the resulting digital pipeline relies solely on the spatial geometry and physical architecture of the hospital buildings, zero patient data is required.
Those anonymised, highly accurate digital twins then provide the vast scale of data and simulation required to rigorously post-train and validate the next-generation VLAs before they interact with a live hospital. An edge case learned in London improves a robot in Glasgow. With an estates portfolio roughly 10x the size of the City of London, the NHS is the perfect environment to hone this pipeline at scale.
Ultimately, this project is helping to ensure that the UK develops sovereign capabilities in critical technologies, with the foundational infrastructure for healthcare robotics built safely within, and for, the NHS, but with global applicability and reach.
We solved the clinical mile.
The next challenge is the clinical metre.
Every sample carried by a robot is time returned to a clinician. Physical AI gives us a path to take the legwork out of logistics and return healthcare professionals to the work that only humans can do.
Written by Tom Watkins
