Robotic bedside surgical assistants: a distinct class of surgical robotics

The bedside-assistant device family

Robotic bedside assistants are best defined first by what unites them as a family and only then by what divides them into meaningful sub-types. The unifying feature is deceptively simple: the device occupies the operative field alongside the surgeon and assumes a supporting task that would otherwise fall to a human assistant, while the surgeon continues to perform the operation. Four characteristics follow from this and together demarcate the family from adjacent technologies.

First, operative continuity: the surgeon continues to perform the principal operative manoeuvres directly, using conventional or substantially conventional minimally invasive instruments, while the system supports a bounded assistant task. This does not require identical ports or wholly standard equipment; it requires that the core operative work remain in the surgeon’s own hands rather than being transferred to proprietary articulated instruments under indirect control. Second, co-located interaction: the robot and the surgeon share a single physical workspace at the table, in contact with the patient - the surgeon is not relocated to a separate console. Third, task-bounded assistance: the device supports one delimited function - endoscope control, retraction, organ positioning - rather than mediating the whole procedure. Fourth, surgeon primacy: the operating surgeon remains the primary operator and decision-maker; the device executes or stabilizes a task within the limits the surgeon sets and can override at any moment.

These criteria draw a clean boundary against telemanipulation, which breaks operative continuity and co-located interaction at once. They also exclude passive mechanical holders with no actuation or control system: these assist positioning but do not participate in the operative task in any responsive sense, and are treated here as non-actuated assistance, outside the actuated family rather than as a zero-capability member of it. What remains inside the boundary is a large and heterogeneous family - three decades of scope holders, manipulators, solo-surgery systems, and co-manipulative assistants - and it is this heterogeneity that the remainder of this section resolves.

Three modes of bedside assistance

The family divides according to how the supporting task is controlled - specifically, where each intraoperative action originates. This is a more reproducible basis for classification than “autonomy”, a term that invites disagreement about thresholds. The operative question is concrete: for any given movement of the device, what determined it - a direct surgeon command, a predefined rule triggered by a set condition, or a selection the system made from its reading of the current context? Three modes follow, and we introduce the shorthand C1-C3 used throughout the remainder of this review.

C1 - surgeon-commanded bedside assistance. Every action of the device is directly and contemporaneously commanded by the surgeon, through a voice interface, head-tracking, a foot pedal, or a handle. The device supplies steadiness and eliminates a human assistant’s tremor and drift, but originates no action of its own. The great majority of the historical camera-holder literature describes systems of this kind.

Delegated rule-based execution (C2) - automation-assisted bedside collaboration. The device executes predefined, rule-based behaviours triggered by specified conditions: recentring the endoscope on a marked instrument, returning to a stored viewpoint, or maintaining a fixed spatial relationship to a tracked target. Action here is not contemporaneously commanded for each movement - distinguishing this mode from C1 - but neither is it independently chosen - distinguishing it from C3. Much of what is marketed as intelligent or AI-enabled camera control belongs here: instrument-following and automated recentring are rule executions over a tracked feature, not context-dependent decisions, unless it is demonstrated that the system is selecting among possible actions rather than applying a fixed mapping.

Context-responsive bounded action selection (C3) - partially autonomous robotic assistance. The device selects an action from its interpretation of the current operative context, rather than executing a single predefined rule. The surgeon supervises continuously and retains override, and the autonomy remains narrow and bounded to the assistant task.

These modes correspond closely to an autonomy-based reading - C1 maintaining a surgeon-set position or executing contemporaneous commands, C2 moving the endoscope or instrument automatically in response to a defined trigger, and C3 selecting an action from the system’s own interpretation of context - but they are defined here by the origin of each action rather than by autonomy level, because autonomy invites disagreement about thresholds whereas the origin of a given movement is comparatively concrete and recoverable from a device’s documentation.

The dividing line the field most often blurs runs between C2 and C3. The operative test is whether the system merely executes a predefined rule over a perceived feature, or independently selects an action from perceived context. Only the latter meets the C3 definition, and, as later sections establish, it is a test the current literature rarely reports enough information to perform.

Explicit decision rules make this classification reproducible. A system is classified as C1 when every device movement is contemporaneously commanded by the surgeon (by voice, head-tracking, foot pedal, or handle). It is classified as C2 when the device executes a fixed, predefined mapping triggered by a specified condition or a tracked feature - for example, recentring on a surgeon-tagged instrument - without contemporaneous command for each movement. It is classified as C3 only when published documentation demonstrates that the device selects among alternative actions on the basis of an interpretation of operative context rather than applying a fixed rule; where the available documentation is insufficient to demonstrate context-based selection, the system is classified conservatively as C2. On this last rule, no system in the present review met the evidentiary threshold for C3 in routine clinical use.

AIONS-consistent collaborative (cobotic) surgery: a separate classification

AIONS consistency is a separate classification from control mode, not a synonym for the highest mode. The consensus defines collaborative surgery by three conjoint requirements: co-located interaction, delegation of a bounded task, and partial autonomy in executing it - the device carries out the delegated task without continuous, action-by-action surgeon command^[20]. These requirements cut across the three modes rather than mapping onto the highest one.

C1 does not meet them: nothing is delegated. C3 clearly meets them. The consequential case is C2. Some C2 systems meet the AIONS lower boundary for collaborative surgery: if a device executes a delegated task over a sustained interval without the surgeon commanding each movement, a bounded task has genuinely been handed over. Others do not: a system that repositions on a single triggered condition and reverts immediately to surgeon command has not been delegated a task in any sustained sense. Whether a given C2 device meets the AIONS lower boundary is an empirical question about how much bounded execution is actually delegated - one the published literature frequently does not provide enough detail to answer, and one on which this review’s own classifications, where offered, are its interpretive application of the AIONS criteria rather than determinations made by the consensus itself.

The core relationship, and a caution against teleology

Bedside assistance is the broad device family. Co-located interaction is necessary but not sufficient for collaborative surgery in the AIONS sense: a device can share the surgeon’s workspace and still be no more than C1. What collaborative surgery additionally requires is partial autonomy together with shared, bounded task execution.

It would be easy, and mistaken, to read C1→C2→C3 as a ladder every device is destined to climb. Partial autonomy is one emerging pathway for the family, visible in a small number of recent platforms; it is not a trajectory the whole family is bound to follow, nor a hierarchy of clinical value. A deliberately surgeon-commanded system may remain the right choice for a given procedure, team, or setting, and may retain clinical value indefinitely even as other devices acquire context-responsive capabilities. The modes classify how a device works, not how advanced or desirable it is.

Positioning the class: two axes

Locating the family requires two conceptual axes, because the two frameworks the field already possesses each capture only one. LASR (Levels of Autonomy in Surgical Robotics)^[39] arranges surgical robots along a single dimension - task autonomy - from Level 1 to Level 5; it cannot distinguish a console-mediated telemanipulator from a co-located bedside assistant at the same nominal level, since interaction configuration is not one of its dimensions. AIONS^[20] supplies that missing dimension - co-located interaction - but as a definitional criterion rather than a continuous scale.

Placed at right angles, the two ideas define the plane on which the class lives [Figure 1]. The horizontal axis is control and task-execution mode - referred to as execution mode throughout the remainder of this review - spanning C1 through C2 to C3, with LASR levels available beneath it as an approximate reference band rather than a one-to-one mapping. The vertical axis is interaction configuration - from console-mediated separation from the bedside workspace (of which geographically distant remote surgery is one case, not the definition) to a co-located, shared bedside workspace. Telemanipulation occupies the console-mediated band; non-actuated holders sit outside the plotted taxonomy; robotic bedside assistants occupy the co-located band across a range of execution-mode positions. The AIONS-consistent cobotic zone is drawn as an overlay enclosing the C3 region together with the portion of the C2 region where sustained bounded execution is genuinely delegated. A light annotation marks the pathway from C1 toward C2/C3 as one emerging direction, not an inevitable progression, noting that surgeon-commanded systems may retain lasting clinical value [Figure 1].

Why AI changes execution, not classification

That agenda begins with the technology often presented, including by manufacturers, as its own solution to the adoption problem: AI. The introduction of AI into bedside assistance is easily misread as the arrival of a new device category. This review treats that reading as a category error. A bedside assistant is defined by the four family properties established above. None of these is a statement about how the device’s movements are computed. AI acts one level down, on control architecture - on what originates each action within the same bounded task. A C2 device that follows a surgeon-marked instrument by a machine-learned image model is holding and moving the laparoscope for the surgeon exactly as a C1 voice-controlled holder did; what has changed is the mechanism of execution, not the identity of the thing doing it.

This review interprets AI’s effect, accordingly, as movement along the horizontal axis of the taxonomy rather than relocation to a new plane. It can convert C1 into C2. And it is the technology through which C2 may, in some future systems, evolve toward C3 - one possible developmental pathway, not an expected trajectory. In both cases, the device remains a bedside assistant performing a single bounded supporting task under a supervising surgeon; only the origin of the action shifts.

Current state

As established above, almost every commercially available bedside assistant with adequate documentation operates as C1 or C2; no verified C3 function exists in routine clinical use, for the reasons of documentation rather than existence set out in the Technology Landscape section. Where C2 capabilities have reached patients, they have done so in early-phase feasibility and development studies rather than comparative trials, establishing that a delegated automated camera-control function can be used in real procedures - not that it selects contextually, or improves outcomes. The honest characterisation of the current frontier is a system at the C2/C3 boundary, cleared and used as a rule-based follower, with context-responsive selection remaining a described aspiration rather than a documented clinical reality.

Shared autonomy

The concept that governs this frontier is shared autonomy, and it must be distinguished from adjacent terms with which it is often conflated. Automation is the execution of a predefined rule, predictable in principle from the rule and its inputs. Full autonomy is the opposite pole: the system pursues a goal and selects and executes the actions required without a human in the control loop. Shared autonomy occupies neither pole - it denotes an arrangement in which a bounded portion of a task is delegated to the device to execute with some independence, while the surgeon remains continuously in supervisory control, able to direct, correct, or resume the task at any moment. C3 is the form shared autonomy takes when the device’s independence extends to selecting actions from its reading of the scene; it is a species of shared autonomy, not a synonym for it, and not yet verified in routine clinical use.

Shared autonomy maps onto the AIONS definition of collaborative surgery set out earlier in this review^[20], and this review reads that mapping as clarifying why the definition is constructed as it is: shared autonomy is the operational content of its partial-autonomy requirement, bounded by the co-location and delegation requirements defined there. This is why, on the interpretive application of the AIONS consensus proposed in this review, a C2 follower can satisfy the AIONS lower boundary for collaborative surgery while remaining well short of context-responsive selection - an interpretive application, not a classification made by AIONS itself. The framework thus treats autonomy not as a single ascending scale but as a bounded region. Confined to bedside assistance, where the delegated task is visualization or positioning rather than tissue manipulation or operative decision-making, shared autonomy is a comparatively contained proposition - one reason this class is a plausible setting in which to develop it, a point the clinical evidence’s findings on safety and feasibility support rather than establish.

Human factors

It is tempting to assume that as a device assumes more of its task, the demands on the surgeon diminish. The opposite is true: increasing a device’s autonomy increases, rather than reduces, the importance of human-factors engineering. Once execution is delegated, the surgeon must maintain an accurate model of a system that now acts partly on its own, and every property governing the quality of that model becomes safety-relevant.

Trust must be calibrated rather than maximal: a surgeon who trusts a follower too little will override it constantly and lose the benefit, while one who trusts it too much may cease to monitor the field it is responsible for keeping in view. Calibrated trust depends on transparency and predictability. A fixed-model C2 follower is predictable in a way a C3 system would not automatically be, which is one reason the C2→C3 transition raises the human-factors stakes rather than settling them.

Two further properties become critical as autonomy grows. Takeover - the surgeon’s ability to resume manual control smoothly and immediately - is the mechanism that keeps shared autonomy shared. Situational awareness must be actively preserved, because delegation carries a characteristic risk: a surgeon relieved of controlling the view may attend to it less, precisely when continued attention is what makes the delegation safe. Cognitive workload is better understood as redistributed than reduced - the surgeon exchanges manual effort for supervisory load, and whether the net effect helps or harms depends on interaction design. Underlying all of this is responsibility, framed here strictly as a matter of supervisory control and system design: the surgeon remains the responsible operator, which is coherent only if the surgeon retains genuine, exercisable supervisory control. These are conditions of the technology’s safe use, not refinements to be added once it matures, and the verified literature’s near-total silence on them is a gap this review’s agenda addresses directly.

Evaluation challenges

The movement of bedside assistants along the execution axis also unsettles the assumptions on which their clinical evaluation has rested. A mechanical holder is a fixed artefact: the device evaluated in a trial is the device deployed afterward. A software-enabled function is not fixed in this way - its behaviour is determined by an algorithm that can be revised, and a device whose behaviour can change is a device whose evidence can expire. As documented above, ScoPilot was cleared alongside a predetermined change control plan permitting pre-specified modifications, an arrangement that explicitly anticipates change within bounds. Such mechanisms are a reasonable response to the reality of software, but they relocate part of the evaluative burden from before deployment to during it - toward ongoing algorithm validation, post-market surveillance, and reproducibility, since a result obtained with one model version may not hold for another.

For the surgeon and the institution, knowing a function was cleared is no longer sufficient to know how it will behave in their operating room. This places a premium on reporting the current literature does not provide: not merely that a device was used and with what outcome, but what function was engaged, in which mode, under which model version, and with what provision for change. The evaluation problem is not primarily regulatory; it is a problem of describing clearly enough what was done that another surgeon could understand and reproduce it. The resulting challenge is therefore not simply to evaluate a device but to evaluate a software-enabled clinical function whose behaviour may evolve over time.

Transition

Bedside assistants are in the middle of a transition in kind of object, even as their class remains constant: from purely mechanical holders, whose evaluation a single study could settle, toward software-enabled collaborative systems whose behaviour is delegated, partly autonomous, and revisable. The taxonomy accommodates this transition without strain, but the methods used to evaluate, report, and adopt these devices have not kept pace with it.

The measurement gap as the field’s central methodological problem

The preceding sections converge on a single diagnosis: a mismatch between the outcomes these devices are intended to influence and the outcomes that studies have measured. Bedside assistants are justified in terms of visualization stability, workflow, ergonomics, staffing, and the safe delegation of a task to software, yet the verified literature measures operative time, and little else, with any consistency. This review interprets that mismatch as the field’s central methodological problem: it explains why three decades of devices have produced so little actionable evidence, and it defines what an adequate evaluation framework must capture.

Workflow and the reallocation of tasks

The most consistently reported observation in the evidence base - reduced dependence on continuous human camera manipulation - is a workflow effect, and yet workflow is among the least rigorously measured dimensions of these devices. Camera control is only one of the bedside assistant’s functions, and delegating it redistributes rather than eliminates the remainder: retraction, exposure, instrument exchange, specimen handling. A device that keeps the view stable but cannot retract has changed what a second person does, not removed the need for one. Workflow evaluation for this class must measure the reorganisation of the whole operative team - flow and disruption, transitions between delegated and manual control, setup demands, and points of physical interference. The verified literature documents such friction only anecdotally; that absence is a defining weakness of the field rather than an incidental one.

Ergonomics, cognitive workload, and situational awareness

Ergonomic effect is one of the few dimensions assessed with validated instruments, and the finding is specific: where measured, it accrued to the assistant rather than the operating surgeon, whose own posture and workload were largely unchanged^[37]. This review reads that result as a caution against assuming that relieving the surgeon of manual camera control necessarily lightens the surgeon’s load. Whether the exchange from manual to supervisory effort is favourable is an empirical question that has not been asked of these systems, because workload has been measured, when at all, as the assistant’s subjective burden rather than the surgeon’s supervisory load, and situational awareness has not been measured at all. Their absence means that the properties on which the safety of shared autonomy depends - trust calibration, predictability, takeover quality, preserved situational awareness - are at present unquantified for every device in this class, and an adequate evaluation framework must treat them as explicitly specified, prospectively measured outcomes rather than secondary refinements.

Economic and health-system value

No study in the verified evidence base includes a formal economic evaluation, and this review regards that as a decisive limitation rather than a minor omission, since the case for these devices is frequently economic. The claim that a device pays for itself by freeing personnel is testable but untested; it rests on assumptions about redeployment, local staffing and wage structures, and disposable and capital costs that vary by institution. The setting in which the economic logic has been made most explicit^[23] - a single-arm holder introduced where camera-person shortages were impairing surgical access - illustrates that health-system value is context-dependent. Credible economic evaluation must specify the staffing model against which savings are claimed, account for setup and turnover time, incorporate the capital risk of platform discontinuation, and relate cost to a measured outcome rather than to time alone.

Toward a framework that measures what these devices are for

Workflow reorganisation, the surgeon’s supervisory workload and situational awareness, the human-factors properties that make delegation safe, and context-specified economic value are precisely the outcomes that would tell a surgeon, an institution, or a regulator whether a bedside assistant is worth adopting - and precisely the outcomes the existing literature has not captured. This review interprets the findings of the preceding sections as establishing that the class cannot be adequately evaluated by perioperative endpoints alone, and that this deficiency becomes more acute as devices acquire behaviour that may itself evolve over time. What follows is a positive obligation: to specify an evaluation pathway, a minimum reporting standard, and a prioritised research agenda - the subject of the next section.

A minimum reporting set

This review proposes that any clinical report of a bedside assistant record, at minimum: the exact device and generation; the software version where a function is software-enabled; the execution mode engaged (C1/C2/C3), specified per operative phase; the bounded supporting task; the operating team’s composition and the surgeon’s experience with the specific system; the setup and port workflow; and a structured account of conversions to open surgery, conversions to human camera control, device failures, instrument-arm collisions, repositioning frequency, and human takeover events.

Each variable earns its place by reference to a demonstrated failure of the current literature. Execution mode and bounded supporting task are required because a device’s C1/C2/C3 status cannot otherwise be recovered and is routinely obscured by marketing language. Software version is required because a machine-learned function can change, and the same product name may denote different behaviour over time. Team composition, surgeon experience, and setup workflow are required because operative-time findings proved procedure-, team-, and learning-curve-dependent. And the record of conversions, failures, collisions, and takeovers is required because these events determine whether delegation was safe and workable, and were documented only anecdotally where at all.

An outcome hierarchy

This review proposes organizing evaluation into three tiers. The primary tier is patient outcomes and safety - complications, conversions, oncologic adequacy, and other patient-centred endpoints. The secondary tier comprises workflow disruption, staffing and task reallocation, ergonomics, supervisory cognitive workload, and situational awareness. The exploratory tier gathers trust calibration, takeover quality, transparency and predictability, economic value, and implementation metrics. Operative time cannot anchor this hierarchy because it pointed in three directions across the reviewed studies and, in the largest randomized trial in the reviewed evidence base, ran against the device; it captures neither primary-tier safety nor secondary-tier workflow and human-factors value.

Maturing the evidence through staged design

Evaluation should progress, following the IDEAL framework for surgical innovation^[51,52], from feasibility and early development^[17-19,50], through comparative cohorts assessed against the outcome hierarchy, to adequately powered multicentre trials with patient-centred endpoints, and onward into implementation studies. The verified literature sits almost entirely at the first stage; the comparative tier consists of fourteen small, heterogeneous, mostly single-centre studies underpowered for patient outcomes. The field’s task is to move deliberately up this ladder, stratifying comparison by execution mode so a C2 function is not pooled with a C1 holder as though they were the same intervention. For software-enabled functions this progression extends into post-market surveillance - a clinical necessity, not a regulatory formality.

Evaluating a function that can change

This review recommends that studies of software-enabled functions report algorithm identity, the software version under study, its update status, and explicitly whether behaviour changed during the study, alongside execution mode. ScoPilot’s authorized predetermined change control plan gives this concrete form: a study conducted before and after such an authorized modification may be studying two different functions under one product name, and recording version and change status is what allows a later reader to know which function a given result describes.

Research priorities

Ten specific priorities follow, each emerging from a demonstrated gap: validated metrics of workflow disruption and operative flow; measurement of the operating surgeon’s supervisory cognitive workload under delegation; measurement of situational awareness under delegation; systematic study of takeover behaviour and its quality; investigation of trust calibration and its determinants in transparency and predictability; context-specified cost-effectiveness analysis tied to a measured outcome; multicentre registries built on the minimum reporting set; adequately powered comparative trials with patient-centred endpoints, stratified by execution mode; comparative evaluation of C2 camera control specifically, the frontier where contemporary evidence is thinnest and clinical claims strongest; and longitudinal post-deployment surveillance linking software versions to outcomes.

Toward the discussion

Bedside assistants have evolved faster than the methods used to evaluate them. A class that began as fixed mechanical holders and is moving, in a small but real vanguard, toward delegated and potentially context-responsive execution has outrun a body of evidence still built around operative time and single-centre feasibility. The remedy is not, in the first instance, better devices, but better evidence: standardized in its reporting, hierarchical in its outcomes, staged in its design, honest about the mutability of software, and directed at the priorities above.

Principal findings

This review asked whether robotic bedside assistants constitute a distinct, clinically meaningful class, and whether the evidence supports their positioning as a step toward AI-enabled surgery. The answer is qualified rather than sweeping. Bedside assistants are best defined by physical relationship and task structure rather than hardware generation. Control and task-execution mode, not device generation or marketing description, is the principal framework by which this class should be classified.

Interpretation: why the field has struggled to see this clearly

The field has generally treated bedside assistants as a single, comparatively simple thing - a camera holder, distinguished from its predecessors mainly by interface - rather than as a family differing in what matters clinically: what the device does with the task once delegated. Kraft’s finding^[35] that an early voice-controlled system was slower and less reliable than a human assistant, and ScoPilot’s clearance as a function executing a bounded task without continuous command, are not two points on a single continuum of “robotic camera assistance”; they describe devices with materially different relationships to the surgeon’s control, and any synthesis that does not first separate them by that relationship will tend to wash out the distinctions on which clinical and regulatory judgment should turn.

Existing frameworks capture complementary but incomplete dimensions of the problem. AIONS^[20] defines collaborative surgery by co-location and partial autonomy; the LASR literature^[39] scales devices by decision-making independence. What was missing was a framework holding the interaction-configuration dimension and the execution-mode dimension together, so that co-location did not have to stand in, by default, for collaboration in the fuller sense the field increasingly cares about. Organizing the evidence this way is what allowed task reallocation to emerge as the field’s best-supported finding and operative time as its least reliable one - a conclusion available in the primary literature all along, but obscured while devices of different execution modes were read as one undifferentiated group.

Clinical implications

For surgeons, a device’s execution mode, not its marketing description, should govern expectations. A C1 holder will do exactly what it is told and no more; a C2 follower will sustain a designated view without further command but will not adapt to context it was not built to recognise. This review’s account of takeover, trust calibration, and situational awareness offers a practical vocabulary for that judgment.

For hospitals and procurement decision-makers, the economic and staffing case for these devices, while plausible, has not been demonstrated in a form that transfers reliably between institutions. The most consistently reported workflow effect - a reduction in the personnel required for camera control - is, on current evidence, modest and incompletely characterised; it does not account for the workflow reorganisation and redistributed tasks that accompany it.

For researchers, the implication is architectural before it is empirical: future comparative work should stratify by execution mode as routinely as trials stratify by disease stage or surgical approach, since pooling across modes reproduces the interpretive difficulty this review set out to resolve.

For regulators, the implication concerns the durability of evidence rather than the adequacy of any particular clearance pathway. As noted above, a predetermined change control plan is a sensible response to software that can change after clearance, but it sharpens rather than resolves the question of how performance is tracked once a function has been modified in the field.

For industry, regulatory clearance of a control architecture, however clearly documented, is not evidence of clinical benefit, and the two claims should not be conflated in how a cleared function is described to clinicians.

For surgical training, delegating camera control to a device has consequences the evidence has not yet addressed. The scopist and bedside-assistant roles have traditionally given trainees early operative exposure and an understanding of exposure and visualization; replacing them with a robotic holder may reduce these opportunities, while potentially offering offsetting benefits such as a more standardized visual field and greater scope for supervised independent operating. Whether the net educational effect is favourable is unknown, and prospective study of the training implications of bedside assistance belongs on the research agenda alongside its clinical evaluation.

AI and the likely evolution of the class

The verified record supports C2: a bounded task, handed over for sustained performance, executed according to a fixed mapping. It does not support a demonstrated, routinely deployed instance of C3. That distinction is not a matter of degree awaiting more capable versions of the same thing; it is a distinction in kind between executing a rule and selecting a response, and the field’s vocabulary should track it precisely rather than allow “AI-enabled” to blur the two. Future development should be evaluated by demonstrated function, not by the label attached to it - a discipline for recognising context-responsive assistance accurately if and when it arrives, and for resisting its premature attribution to systems that have not yet demonstrated it.

Strengths of this review

The review’s principal strength is methodological: every device classification rests on primary regulatory documentation or peer-reviewed technical description rather than manufacturer characterisation where such documentation was available, and every clinical claim was rebuilt from the full text of the underlying studies rather than inherited from prior syntheses. This process surfaced eligibility errors - two studies (Uchal et al.^[48] and Kondraske et al.^[49]) initially identified as candidate clinical evidence but excluded on full-text review as, respectively, a simulator study and an animal study; a comparative arm found to combine active and passive devices^[21] - that a less granular approach would have carried forward silently. The organisation of the technology and evidence around execution mode is this review’s central conceptual contribution.

Limitations

The field is changing quickly, and any account of the technology landscape risks being overtaken by subsequent clearances or discontinuations. The comparative clinical evidence, even rebuilt from primary sources, remains limited to a small number of heterogeneous, mostly single-centre studies, several confounded by non-random allocation, insufficient for quantitative synthesis and reported here in narrative form only. Device classification depended on publicly available regulatory summaries, which for several systems proved unavailable in sufficient detail, so a number of classifications are explicitly marked as unverified rather than assigned by inference. And the heterogeneity that motivated this review’s organizing framework is also, unavoidably, a limitation of the review itself: it precludes any pooled estimate of effect.

Concluding perspective

The trajectory traced across this review - from C1 holders, through C2 execution, toward a still-unverified prospect of C3 assistance - describes a class of devices whose engineering has, in places, moved ahead of the evidence and methods used to understand it. The challenge that now defines the field is accordingly less about building more capable bedside assistants than about learning to evaluate, report, compare, and safely integrate the software-enabled collaborative functions already emerging.

This review’s lasting contribution is not a verdict on any single device or platform. It is the demonstration that control and task-execution mode - not hardware generation, not marketing language, not autonomy level alone - is the framework by which this class should be classified, evaluated, and reported, together with an evidence base rebuilt from primary sources and a reporting and research agenda specified in enough detail to be followed. If bedside assistants are indeed the substrate on which shared autonomy and AI-enabled surgery will first be validated, the execution-mode framework proposed here is offered as the discipline that translation will need: a way of ensuring that what gets called collaborative, or intelligent, or autonomous, has actually earned that description.