From Report Cards to Co-Pilots: The Technical Reality of AI in Robotic Surgery

Maya: Here is the reality check for the future of robotic surgery: out of all the recent studies trying to build AI that actually collaborates with a surgeon in the operating room, 96.7% of them provide absolutely zero autonomous functionality.

Sam: That is a staggering drop-off from the hype we usually hear about surgical AI. Today we are unpacking a scoping review titled Machine Learning for Human-Autonomy Teaming in Surgical Skill Assessment, published in JMIR AI.

Maya: And before we get into why the jump from post-op report cards to live robotic co-pilots is so extraordinarily difficult, a quick reminder that our voices are AI-generated, and this is a summary of a public document.

Sam: So, Maya, the document focuses on something called human-autonomy teaming, or HAT. That term feels like a paradigm shift. We are not just talking about a tool that scores a resident's suturing after the fact.

Maya: Exactly. Human-autonomy teaming represents a shift away from traditional automation, where a machine operates either in total isolation or under strict human control. HAT involves dynamic cooperation. The human and the AI share decision-making, situational awareness, and task execution.

Sam: Like a true partnership in a high-stakes clinical environment. The AI would analyze the surgical context in real time, adjust to the surgeon's skill level, and provide tailored feedback on the fly. But based on your opening stat, we are not there yet.

Maya: We are very far from it. The researchers systematically searched PubMed, IEEE Xplore, and Web of Science for peer-reviewed studies published between 2019 and 2025. They identified 92 studies that applied machine learning to surgical skill assessment.

Sam: And what did they find when they evaluated the technical maturity of these 92 studies? The paper mentions using the NASA technology readiness level scale, or TRL, which goes from basic research at TRL 1 up to proven operational deployment at TRL 9.

Maya: The TRL analysis is incredibly telling. They found that 77.2% of the studies—that is 71 out of 92—remain at TRL 3. That means they are just proof-of-concept models tested on limited or benchmark data. Only 2.2%, or 2 out of the 92 studies, reached TRL 5 or above, which involves testing on clinical or cross-institutional data.

Sam: Wait, really? So much of the research is stuck at the proof-of-concept phase? That seems like a massive bottleneck for any medtech company trying to commercialize this stuff.

Maya: It is a massive bottleneck. And it gets even more stark when you look at the Yang levels of autonomy for surgical robotics. The Yang framework defines 6 stages. Level 0 is no autonomy, pure teleoperation. Level 1 is robot assistance, like tremor filtering. Level 2 is task autonomy for specific subtasks. Level 3 is conditional autonomy. And levels 4 and 5 are high and full autonomy.

Sam: So if human-autonomy teaming requires the AI to actively support decision-making, we need systems operating at least at Yang level 2 or 3, right?

Maya: Right. But the authors mapped the literature to this framework and found that 89 out of the 92 studies—96.7%—operate at Yang autonomy level 0. They provide post hoc classification or passive feedback, but zero autonomous functionality. Only 3 studies reached level 1. Absolutely none demonstrated level 2 or higher.

Sam: That completely reshapes how I view the current landscape. We hear so much about AI in surgery, but we are essentially still just building highly sophisticated assessment tools, not robotic teammates.

Maya: Exactly. The research proves we have a robust foundation for objective skill assessment. Traditionally, evaluating a surgeon's technical skills like instrument handling was done through expert observation using rating tools like OSATS or the Global Rating Scale. Those are time-consuming and subjective. Machine learning offers scalable, objective, data-driven assessment. But moving from assessment to teaming requires overcoming huge technical hurdles.

Sam: Let's dig into those hurdles. Where is the data coming from to train these AI models? Because I imagine teaching a machine to understand a cholecystectomy or a robotic prostatectomy takes an enormous amount of varied data.

Maya: That brings us to the first major limitation: ecological validity. A massive chunk of foundational work relies on one specific benchmark dataset called JIGSAWS. It stands for the Johns Hopkins University and Intuitive Surgical Inc Gesture and Skill Assessment Working Set.

Sam: I have heard of JIGSAWS. It sounds very official, but what is actually in it?

Maya: It includes synchronized video and kinematic data of simulated robotic tasks, like suturing and knot tying. But here is the catch: it only has 39 trials performed by just 8 participants in a highly controlled simulation environment.

Sam: 8 participants? You cannot possibly capture the complexity of real-world anatomy and unpredictable operating room environments from 8 people doing simulated knots.

Maya: You absolutely cannot. And that is why the high performance reported on these benchmark datasets can be misleading. The deep learning architectures might just be memorizing dataset-specific patterns rather than learning transferable surgical skills. When models trained on data like JIGSAWS are applied to real clinical settings, they often struggle to generalize.

Sam: The domain gap between a simulator and a real patient is huge. How many of these 92 studies actually tested their AI outside of the original dataset?

Maya: Not enough. The review found that 38% of the studies—35 out of 92—lacked any form of external validation whatsoever. And only 15.2%, or 14 studies, conducted cross-site testing with data from a different institution.

Sam: So if you are a health-system leader looking to buy one of these AI tools, you have to ask: was this validated on diverse clinical data, or just a small, homogenous lab set? Because if it is the latter, it might fall apart in your specific operating rooms.

Maya: Precisely. Now, to be fair, the field is shifting. The review found an even split overall: 46 of the papers used simulation-based setups, and the other 46 focused on clinical or semiclinical data, including live intraoperative recordings and cadaver models. We are seeing models trained on real laparoscopic gastrectomies, cataract surgeries, and colorectal procedures.

Sam: Let's talk about how the AI is actually watching these procedures. What kind of data modalities are they feeding into the machine learning models?

Maya: Video is the dominant modality. Out of the 92 studies, 58 used video data, primarily analyzing endoscopic or external camera footage. Another 18 relied strictly on kinematic data, which is essentially the robotic system's sensor data capturing tool tip positions, velocities, and forces.

Sam: Video makes sense because it is already being recorded in so many laparoscopic and robotic surgeries. It doesn't disrupt the workflow. But I imagine video can be messy. You have variable lighting, blood, camera movement, instruments blocking the view.

Maya: Exactly, which is why 12 of the studies used multimodal approaches. They fused data streams—like combining video with kinematics, or even physiological data like electroencephalography, to create more robust representations of the surgical performance.

Sam: Wait, electroencephalography? They are putting sensors on surgeons' heads to measure brain activity during surgery?

Maya: Quick note before we carry on. This spot is open for a sponsor. If your company builds or sells AI for healthcare and wants to reach the clinicians, health-system leaders and industry teams who listen to this show, the link to our sponsorship page is in the show notes.

Sam: And now, back to the document.

Maya: Yes, 3 studies looked at physiological signals. One study combined EEG and eye-tracking in live robotic setups to detect workload and subtask transitions in real time. It is early-stage, but it shows how AI could infer a surgeon's cognitive load and motor intention, which is a critical component for adaptive feedback.

Sam: That is wild. If the system knows the surgeon is experiencing high cognitive load during a complex dissection, it could theoretically adapt its feedback or suppress non-urgent alerts. But what kind of machine learning is actually crunching all this video, motion, and brain wave data?

Maya: Deep learning has completely taken over. Convolutional neural networks, or CNNs, were the most common technique, appearing in 55 of the studies. They are used for everything from object detection to extracting spatial patterns from video frames. Recurrent models like LSTMs were used in 14 studies to capture temporal dependencies—basically understanding the sequence of surgical gestures over time.

Sam: Because surgery isn't just a static image. It is a sequence of actions. You have to know what just happened to understand what is happening now.

Maya: Exactly. And recently, there is a shift toward transformer-based architectures, which were used in 11 studies. Transformers are great at capturing long-range dependencies across continuous video streams. There is also a push into graph-based learning. For example, some models use graph convolutional networks to treat video features and left-hand and right-hand kinematics as distinct nodes in a graph, mapping relationships like vision-to-motion and in-between-motions.

Sam: Okay, so the underlying architecture is getting incredibly sophisticated. But for human-autonomy teaming to work, this processing has to happen instantly. A co-pilot isn't helpful if it tells you about a mistake seconds after you made it. How do these models perform on latency?

Maya: Latency is a massive hurdle. Out of all the studies, only 25%—that is 23 out of 92—demonstrated real-time capability. The majority, 53 studies, performed inference purely offline, retrospectively. And 16 were classified as near-real-time.

Sam: So the vast majority of this research isn't even functioning in real time yet.

Maya: Right. The authors highlight an engineering tension between model complexity and latency. To get high reliability, you need deep, complex models. But running those requires serious GPU infrastructure, which most operating rooms simply do not have. So developers either have to compress their models, which risks lowering accuracy, or rely on edge computing architectures optimized for the OR.

Sam: That is a huge takeaway for digital health companies. Building the algorithm is only half the battle; packaging it to run on hospital hardware with millisecond latency without compromising safety is the real product challenge. What about the feedback itself? How does the AI explain its decisions to the surgeon?

Maya: That brings up the issue of interpretability. The review found a disconnect between what ML researchers consider interpretability and what surgeons actually need. Surgeons need actionable, real-time justifications embedded in their clinical workflows. They need decision-centric explainability.

Sam: Instead of what? What are the researchers currently building?

Maya: Currently, 71.7% of the studies report some form of built-in interpretability, but that largely just means using attention mechanisms or feature importance rankings. Only 5.4% used dedicated post hoc explanation methods like SHAP or Grad-CAM. The review argues that rather than just appending post hoc tools, future systems need inherently transparent modules—mapping internal data to clinically meaningful features like tissue tension or proximity to critical structures.

Sam: That makes total sense. A surgeon doesn't want a heat map of pixels; they want the system to say, 'I am flagging this because the tissue tension is abnormally high.' They need the system to communicate uncertainty, too. If the AI is confused by blood in the camera view, it should defer to the human, not guess.

Maya: Exactly. True teaming requires the AI to communicate confidence levels. But another massive gap in current readiness is adaptivity. Surgery is deeply personal. Different experts have different techniques and pacing. Yet, 91.3% of the reviewed models—84 out of 92—are static, with no user-specific adjustment.

Sam: Wait, really? The models assume every surgeon operates exactly the same way?

Maya: Yes. Only 6.5%, or 6 studies, demonstrated any form of adaptive behavior. The root cause is data scarcity. To adapt to a specific surgeon, the model needs sufficient per-user data, which is rarely available. The authors suggest future research needs to look into meta-learning frameworks that allow for rapid few-shot adaptation based on minimal calibration trials, or online continual learning where the model updates as per-user data accumulates.

Sam: I can see the regulatory bodies having a field day with an AI model that learns and changes its parameters during a live surgery.

Maya: You hit the nail on the head. The review explicitly calls out the regulatory and medico-legal considerations that the literature largely overlooks. AI systems that influence intraoperative decisions will require regulatory clearance through strict pathways like the US Food and Drug Administration's Software as a Medical Device framework, or the EU Medical Device Regulation.

Sam: And from a legal standpoint, who gets sued when a robotic co-pilot makes a bad call? If the AI recommends a dissection plane, the surgeon follows it, and there is an adverse outcome, how do you allocate liability among the surgeon, the hospital, and the algorithm developer?

Maya: The document states that the allocation of responsibility remains legally ambiguous in most jurisdictions. These uncertainties create institutional risk aversion. Even if the technology was perfectly ready today, hospitals might hesitate to adopt it simply because the legal framework hasn't caught up.

Sam: So if we synthesize all of this for the people listening—the medtech executives, the health-system leaders—what is the immediate path forward? How do we get out of this TRL 3, Yang level 0 rut?

Maya: The authors lay out several concrete research priorities. First, we have to solve the data bottleneck through coordinated, multicenter dataset initiatives. They suggest federated learning, where institutions collaboratively train models on distributed clinical data without sharing sensitive patient recordings. This sidesteps major privacy barriers.

Sam: That would instantly improve the ecological validity and help models generalize across different hospitals and patient demographics.

Maya: Second, we need to move from generic assessment to individualized decision augmentation. Future models must integrate patient-specific context—like preoperative imaging and pathology-specific risk profiles—so the AI can anticipate procedure-specific challenges.

Sam: Right, a surgery isn't happening in a vacuum. The AI needs to know if the patient has anomalous anatomy before the first incision.

Maya: And finally, we have to advance from state recognition to procedural policy learning. It is not enough for the AI to recognize what gesture is currently happening; it needs to learn the underlying grammar of the surgical task so it can recommend what should happen next. That is the foundational requirement for actual decision support.

Sam: This scoping review is such a necessary reality check. It cuts through the marketing language of AI surgery and shows that while we have brilliant tools for objective skill assessment—like identifying whether a trainee's suturing motions are smooth—we are still building the basic scaffolding for true human-autonomy teaming.

Maya: It is a sociotechnical challenge as much as a computational one. For developers, the message is clear: stop relying solely on benchmark datasets like JIGSAWS, prioritize real-time edge computing, and codevelop interfaces with surgical teams. For clinicians, expect highly intelligent assessment tools in the near term, but know that a true robotic co-pilot is going to take time, massive multicenter data collaboration, and regulatory clarity.

Sam: That is an excellent summary. For those who want to read the full study, we have linked the source document in the show notes. And as a final reminder, this podcast is just a summary of a public document and is not medical advice.