Surgical Robotics and AI: What We Achieved in 2025-2026, and What We Have Not Yet Proven

For more than two decades, surgical robotics evolved largely in the shadow of a single dominant player. Yet 2024-2025 marked an inflection point on several fronts at once: a new generation of platforms won regulatory clearance, the tactile feedback missing for decades was finally productized, computer vision began to filter into the operating room, and autonomous surgery completed, for the first time in a laboratory, a surgical phase that was both learned and correctable by a surgeon's voice. This article summarizes the current evidence with deliberate restraint, separating what has been proven from what has not yet been proven. Where data conflict, the findings are presented side by side rather than declaring any single source definitive.

The competitive landscape has shifted: three new platforms on stage

The sector's most consequential news is not clinical but structural: Intuitive's roughly twenty-year de facto monopoly in the multi-port market has been broken. The Medtronic Hugo RAS system received FDA clearance in the United States on December 3, 2025 for urologic procedures—prostatectomy, nephrectomy, and cystectomy—representing a market of roughly 230,000 surgeries per year. The supporting Expand URO trial was presented as the largest completed IDE study for multi-port robotic urology in the U.S. (n=137: 55 prostatectomies, 29 cystectomies, 53 nephrectomies). In the prostatectomy arm, the Grade ≥3 complication rate was reported at 3.7% (95% CI 0.5-12.7; P=0.0006), well below the 20% performance goal. Two deaths in the trial were adjudicated by an independent monitoring committee as unrelated to the system.

In the same window, the CMR Surgical Versius Plus received 510(k) clearance in December 2025 for cholecystectomy; with its modular design and lower cost (~USD 1-1.5 million), it is seen as reinforcing a "a robot in every hospital" trend. Alongside these two competitors, Intuitive's own flagship da Vinci 5 was cleared by the FDA on March 14, 2024, introducing the first integrated Force Feedback in robotic surgery and staking a claim to close the long-standing tactile-feedback gap.

Force feedback: the measured gain is real, the clinical gain is not yet

A long-recognized weakness of robotic surgery has been the surgeon's inability to "feel" how much force is applied to tissue. The force feedback introduced with da Vinci 5 targets precisely this gap. Preclinical data reported a reduction in applied tissue force of up to 43% across every level of experience. In an independent single-blind randomized preclinical study (Servais et al., Surgical Endoscopy 2024; n=29 novice surgeons), force feedback reduced mean applied force to 1.96 N versus 2.53 N (p=0.005) and mean error to 0.44 versus 0.88 (p=0.012), with significant improvements in suture breakage and tissue-trauma scores as well.

Here honesty is essential: an improvement in a laboratory force metric does not equal superiority in a hard clinical endpoint. In the first peer-reviewed clinical pilot of da Vinci 5 (Rashidi et al., The American Journal of Surgery 2025; 68 colectomies, single experienced surgeon, retrospective), return of bowel function averaged 0.8 days and showed no significant difference with force feedback (p=0.12). Placed side by side, the two findings read as follows: the force reduction is a strong, reproducible signal, but its translation into a complication or recovery advantage has not yet been demonstrated. The single-center, single-surgeon, retrospective design lowers the level of evidence of this pilot; prospective randomized trials that could settle the question (e.g., ClinicalTrials.gov NCT07247175) are still at the registration stage.

Autonomous surgery: a major leap in the lab, zero in humans

The most striking scientific event of 2025 was the SRT-H (Surgical Robot Transformer–Hierarchy) study (Kim et al., Science Robotics, July 2025). The system uses language-conditioned hierarchical imitation learning: a high-level planner generates task and correction instructions in language space, while a low-level policy produces motion trajectories. On ex vivo cholecystectomy it achieved 100% success across 8 different gallbladders without human intervention, and the robot adapted in real time to a surgeon's spoken corrections such as "move a little to the left." What matters is not the completion of a single motion but the first learned, correctable execution of a long-horizon, multi-step surgical phase.

This is the natural evolution of the 2022 STAR line (supervised autonomy in bowel anastomosis): a shift from "a single preprogrammed task" to "a learned, correctable multi-step phase." But the boundary is clear and should not be overstated: the work was performed on porcine tissue, ex vivo. The bleeding of live perfused tissue, respiratory motion, and genuine anatomical variation were not tested; there is as yet no transition to an in vivo or human setting.

Viewed across the whole field, autonomy is still strikingly early. A systematic review by a Mount Sinai team (Lee et al., npj Digital Medicine 2024) classified 49 FDA-cleared surgical robots from 2015-2023 on the LASR scale (0-5):

In other words, the vast majority of cleared robots remain at an "assistive" level operating under continuous surgeon control; only two robots declared formal machine learning in their FDA filings, and only 39% of the systems included clinical test data. No fully autonomous surgical robot exists.

Computer vision: high technical accuracy, limited real-time use

The second pillar of AI in the operating room is computer vision: real-time recognition of the surgical phase, anatomical structures, and instruments. The technical accuracy is impressive—systematic reviews report phase-recognition accuracy of 81-93.2% and anatomical-structure recognition of 71.4-98.1% (Paracchini et al., AOGS 2025). As an example of safety applications, "Go/No-Go" decision-support models are being developed to mark safe and unsafe dissection zones in laparoscopic cholecystectomy with the goal of preventing bile duct injury.

The point of honesty here is that accuracy figures alone do not mean readiness for real-world use. In one scoping review, 490 articles were screened and 113 included, yet only 12% (13 studies) evaluated real-time intraoperative integration; the remainder were retrospective feasibility studies (Art Int Surg 2025). More strikingly, a large systematic review of reporting quality (Carstens et al., medRxiv 2025—188 studies; to be labeled as a pre-peer-review preprint) found that most studies were small, single-center, and frequently dependent on the same cholecystectomy dataset, with limited external-dataset testing, and that clinical translation was addressed in only 11 studies. In short: laboratory accuracy is high, clinical transition is still early.

Regulation and systematic risks

In January 2025 the FDA issued draft guidance for AI-enabled device software, foregrounding transparency, model inputs and outputs, performance and bias disclosure, and the notion of a "predetermined change control plan." This is an important step toward answering how continuously learning systems should be updated after clearance. Even so, the systematic risks facing the technology must be listed honestly:

Distribution shift and atypical anatomy: Systems trained by imitation learning may fail to generalize in rare scenarios underrepresented in the training data or in abnormal anatomy; a system that "works in normal anatomy" is a safety liability in the atypical case. Automation bias: Over-reliance on AI can erode the surgeon's clinical judgment. Data bias and inequity: Imbalanced training data can amplify demographic differences in outcomes, and high-cost platforms can concentrate care in wealthier centers. Publication bias: Positive, industry-sponsored preclinical data dominate while independent randomized trials are scarce—hence the need to monitor ClinicalTrials.gov and PROSPERO registrations. Liability gap: If harm occurs in a fully autonomous system, who bears responsibility is legally unsettled; classification and regulation lag behind technological progress.

Conclusion

2025-2026 is a period of genuine momentum for AI in surgical robotics, but momentum must be separated from outcome. What has been proven is concrete: new platforms (Hugo, Versius Plus) cleared regulatory safety and effectiveness thresholds and broke Intuitive's monopoly; force feedback measurably reduced applied tissue force; an autonomous system completed a surgical phase in the laboratory with high success and in a correctable manner; and computer vision reached high technical accuracy on recognition tasks. What has not yet been proven is equally important: a hard clinical benefit of force feedback has not been shown (no difference in the dV5 pilot, p=0.12); autonomy has not been tested in humans; and the real-time clinical integration of computer vision is very limited (~12%) with weak external validation. The most balanced reading is that AI has matured today to augment rather than replace the surgeon's judgment. The progress is real, and managing expectations is mandatory—because surgery is among the domains where the cost of error is highest.