Autonomous Surgical Robotic Assistants (ASARs)

• ASARs are autonomous robotic systems that actively assist in minimally invasive surgery by automating repetitive or ergonomically demanding tasks under human oversight.
• They employ diverse control paradigms—including symbolic reasoning, visual servoing, and learned multimodal policies—to ensure precise and safe task execution.
• Their integration into surgical workflows minimizes surgeon cognitive load and improves efficiency through validated applications such as endoscope guidance and instrument handling.

Autonomous Surgical Robotic Assistants (ASARs) are active collaborative partners in robot-assisted minimally invasive surgery in which the primary surgeon remains in overall control while a secondary robotic system autonomously performs a supporting function such as camera navigation, tissue holding, tensioning, suction, or complementary manipulation (Colan et al., 15 Jul 2025). Within the broader surgical-autonomy continuum, ASARs are best situated between pure teleoperation and full autonomy: they automate bounded supportive functions, often under surgeon supervision or approval, rather than replacing the surgeon as the primary operative agent (Yip et al., 2017, Singh, 2022). Their practical significance lies in offloading repetitive, ergonomically demanding, or cognitively disruptive subtasks while preserving human judgment over strategy, safety, and exception handling.

1. Conceptual scope and autonomy regimes

ASARs are distinguished from conventional teleoperated robots and from passive guidance systems by their autonomous assistant role. In the review literature, this role is defined specifically for scenarios in which the robot provides meaningful and active support to a human surgeon, rather than functioning only as a slave manipulator or as a shared-control constraint layer without independent assistant behavior (Colan et al., 15 Jul 2025). This framing is consistent with broader taxonomies of surgical autonomy that separate direct control, shared control, supervised autonomy, and full autonomy, and that treat supervised autonomy as especially suitable for surgery because it combines robotic precision with human oversight (Yip et al., 2017).

Two collaborative setups recur across the literature. In teleoperation-based assistance, the lead surgeon operates the primary instruments while the assistant robot controls a complementary function such as endoscope guidance or auxiliary tool manipulation. In direct hands-on interaction, the surgeon and robot coexist in a shared workspace at the bedside, creating tighter coordination demands but also enabling more immediate collaborative behavior (Colan et al., 15 Jul 2025). A related systems-level perspective extends autonomy beyond the operative phase and treats preoperative, operative, and postoperative functions as a staged continuum, with levels ranging from no autonomy to full autonomy and with near-term emphasis on assistive and conditional-autonomy systems (Singh, 2022).

A recurring clarification in the literature is that ASARs are not equivalent to autonomous primary surgeons. The humanoid proof-of-concept study on a cadaveric sphenoidectomy is explicit on this point: it studies a specific class of ASARs, namely robotic first assistants that can occupy a human assistant role inside an existing operating-room workflow, rather than autonomous primary surgeons (Cho et al., 27 Feb 2026). This distinction is foundational because many of the most credible near-term assistant tasks involve visualization, exposure, instrument handling, and workflow support rather than autonomous dissection or full procedure execution.

2. Principal assistant functions and operative roles

Endoscope guidance is the most mature ASAR application. In the systematic review, 69% of included studies addressed endoscope guidance, reflecting the fact that camera management is repetitive, cognitively burdensome, and naturally aligned with an assistant role (Colan et al., 15 Jul 2025). Representative systems include autonomous instrument-tracking endoscope holders that accept semantically meaningful goals such as focusing on the right-hand instrument, the left-hand instrument, or a point between the instruments, thereby shifting camera control from low-level commands to task-level assistant behavior (Belmonte et al., 2021). Gaze-based learning from demonstration extends the same role into robot-assisted surgery on the da Vinci platform by learning ECM positioning from surgeon gaze and kinematics, targeting the mode-switching burden that occurs when the surgeon must manually leave instrument control to reposition the camera (Abdelaal et al., 2023).

Exposure generation and tissue tensioning form the second major cluster. MEDiC automates robotic assistance to maximize visual exposure and apply tissue tension for a specified dissection site, with the human retaining control of the cautery tool while the robot acts as the assistant hand (Liang et al., 2024). A complementary line of work addresses autonomous tissue retraction through interpretable task reasoning and biomechanical simulation, explicitly modeling retraction as a safety-critical assistant task in deformable anatomy rather than as autonomous primary surgery (Meli et al., 2021). Force-aware autonomous retraction on the dVRK further narrows the assistant role to automated tissue lifting, but shows that including tool–tissue interaction forces improves both success rate and gentleness relative to a vision-and-kinematics baseline (Abdelaal et al., 20 Jan 2025).

Instrument handling and scrub-nurse behavior constitute a third ASAR domain. Autonomous pick-up of suturing needles targets the preparatory bottleneck before urethrovesical anastomosis by detecting, approaching, and grasping a needle in a suturing-ready pose that avoids hand-offs (D'Ettorre et al., 2018). More recent robotic scrub-nurse systems expand this into language-conditioned instrument delivery. One dual-arm platform using two Franka Research 3 arms, RGB-D perception, GPT-4o-based task planning, and a unified quadratic-programming safety layer reported an 83.33% success rate in surgical instrument delivery with no collisions in 30 trials (Luo et al., 3 Mar 2026). RoboNurse-VLA, which integrates YOLOv8, SAM 2, and OpenVLA/Llama 2 for voice-commanded handover, reported 100% success on the standard “On table” task, 95% under hand-height or hand-pose changes, 90% on unseen tools, and 95% on difficult-to-grasp items (Li et al., 2024).

The literature also includes collaborative manipulation roles beyond simple holding. Multi-agent reinforcement learning has been used to create hybrid human-AI teams for exposure and dissection primitives in simulated laparoscopic cholecystectomy, where one agent manipulates a gripper and the other a cautery hook, directly modeling the assistant as a cooperative teammate controlling one instrument while the human controls the other (Scheikl et al., 2021). This suggests that ASAR roles need not be limited to static support; they can extend to role-specialized concurrent collaboration.

3. Architectural patterns and control paradigms

ASAR architectures are heterogeneous, but several recurring designs are evident. One family combines symbolic task reasoning with situation awareness and low-level motion generation. In “Autonomous task planning and situation awareness in robotic surgery,” the architecture joins Answer Set Programming for explainable task planning, Dynamic Movement Primitives for dexterous execution, and a situation-awareness module that interprets sensor data, detects anomalies, and triggers replanning (Ginesi et al., 2020). FRAS develops the same logic further for deformable tissue retraction by integrating a logic module for task-level interpretable reasoning, a biomechanical simulation that complements real sensors, and a situation-awareness module for context interpretation (Meli et al., 2021). These systems are notable because they treat explainability and failure recovery as architectural requirements rather than post hoc add-ons.

A second family uses differentiable physical models and visual servoing. MEDiC represents tissue as a volumetric deformable object, estimates state by registering a simulated mesh to stereo observations, and uses deformable Jacobian analysis both to choose assistance contact points and to control motions that increase wedge opening, regulate shear, and enforce local stretch around a surgeon-defined dissection line (Liang et al., 2024). This approach is emblematic of ASARs for deformable anatomy: assistance is formulated not as generic motion generation but as optimization of exposure and tension under tissue mechanics.

A third family centers on learned multimodal policies. The gaze-based camera-assistance work uses Gaussian Mixture Models and Gaussian Mixture Regression to map surgeon gaze, instrument state, and robot kinematics to desired ECM pose, coupled with workspace validation to ensure reachability and non-singularity (Abdelaal et al., 2023). Force-aware retraction uses Action-Chunking Transformers trained by imitation learning, comparing a “force policy” that uses force, vision, and robot kinematics against a “no force policy” using vision and kinematics alone (Abdelaal et al., 20 Jan 2025). Foundation-model-based assistants extend this trend into language-conditioned manipulation: the dual-arm scrub-nurse system uses DINOv2, SAM, Mediapipe, GPT-4o, and a unified quadratic program for task tracking, obstacle avoidance, and self-collision prevention, while RoboNurse-VLA uses SAM 2, a projection module, and a Llama 2 backbone in OpenVLA to map voice and scene observations to tokenized robot actions (Luo et al., 3 Mar 2026, Li et al., 2024).

A fourth architectural concern is the hardware-software substrate itself. An open-source platform for autonomous laparoscopic surgery argues that existing research platforms based primarily on the dVRK suffer from cable-driven compliance, backlash, and hysteresis that degrade state-space consistency, and therefore proposes a robot-agnostic Remote Center of Motion controller with a closed-form analytical velocity solver for industrial manipulators such as the UR5e and Franka Emika Panda (Rodriguez et al., 9 Mar 2026). The platform explicitly supports teleoperation, demonstration recording, and deployment of learned policies, making infrastructure precision itself an ASAR issue rather than a neutral background assumption.

4. Human–robot interaction, intent exchange, and supervision

ASARs depend on how intent is communicated and how autonomy is bounded. A central result of the review literature is that wider adoption is hindered by preference alignment, procedural awareness, skill acquisition, and human–robot information exchange (Colan et al., 15 Jul 2025). In practice, this means the assistant must not only act correctly in a geometric sense, but do so in a way that matches surgeon expectations, procedural timing, and communication conventions.

Several interaction modalities have therefore been explored. Voice control is one direct supervisory layer. A ROS-based interface using Whisper, a mapping module based on Word Error Rate, and a Kinova Gen3 with an OpenRST tool supports seven task-level commands—“hey robot,” “tense,” “release,” “pull more,” “pull less,” “stop,” and “terminate”—with average command latencies from 1.14 s to 2.50 s and an overall average of approximately 1.7 s (Davila et al., 2024). The importance of this result is not full autonomy, but the demonstration that surgeon-robot collaboration can be mediated through discrete, safety-bounded action primitives rather than low-level teleoperation.

Implicit intent inference is another interaction strategy. In the da Vinci camera-assistance study, adding orientation and eye gaze to position information reduced average position error to 0.05 mm with standard deviation 2.0 mm and achieved average orientation error of $0.97^\circ$ with standard deviation $1.14^\circ$, compared with a position-only model that achieved root mean square Cartesian position error of 0.1 mm with standard deviation 4.3 mm (Abdelaal et al., 2023). The evidence is therefore suggestive that gaze can encode intended viewpoint selection, although the study does not isolate the marginal contribution of gaze alone.

Embodied first-assistant systems reveal further HRI complexity. In the humanoid endoscopic assistant study, the cognitive demands of assistance—anatomical perception, anticipation of the surgeon’s next move, interpretation of verbal cues, collision avoidance in context, and responsive adjustment of the view—were supplied entirely by a trained teleoperator rather than by robot autonomy (Cho et al., 27 Feb 2026). The same study reports that surgeon and teleoperator had to establish a common language, particularly because medical instructions are often given relative to the patient, and that practice improved coordination substantially. This indicates that ASAR communication cannot be reduced to generic command vocabularies; patient-relative coordinate conventions and context grounding are integral.

Natural fallback and control transfer are also important. In the autonomous instrument-tracking endoscope system, the robot arm is mechanically backdrivable, allowing the surgeon to physically guide the camera in comanipulation mode when autonomous tracking becomes inadequate (Belmonte et al., 2021). This design reflects a broader ASAR principle: shared autonomy should be interruptible and recoverable without forcing a disruptive interface transition.

5. Empirical evidence, benchmarks, and current evidentiary limits

The empirical record for ASARs is promising but still predominantly preclinical. In the 32-study review, 62% of experimental settings were phantom-based, 19% were simulation, and 19% were ex vivo; most technologies therefore remain in pre-clinical development (Colan et al., 15 Jul 2025). Endoscope guidance has the deepest evidence base, while autonomous tool manipulation, tissue handling, and scrub-nurse tasks are newer and more heterogeneous.

Several studies nevertheless provide concrete quantitative support for assistant autonomy. In hybrid human-AI teams for simulated gallbladder exposure and cautery, completion times were 44.4% to 71.2% shorter and collisions were 44.7% to 98.0% fewer than in human teams, although path lengths increased under artificial control by 11.4% to 33.5% (Scheikl et al., 2021). MEDiC reported an average success rate of 82% over 33 phantom trials, compared with 11% for its open-loop hinge baseline, and achieved an average final expansion ratio of $1.62 \pm 0.41$ versus $1.14 \pm 0.23$ (Liang et al., 2024). Force-aware retraction achieved 76% versus 26% success on a previously seen tissue sample and 70% versus 20% on a previously unseen tissue sample, while exerting lower average forces than the no-force baseline in both settings (Abdelaal et al., 20 Jan 2025). An open-source laparoscopic autonomy platform reported sub-millimeter RCM deviations in phantom, ex vivo, and in vivo porcine settings and an 85% success rate, 17/20 roll-outs, for a learned bowel-grasping-and-retraction policy (Rodriguez et al., 9 Mar 2026).

At the same time, multiple papers emphasize qualitative feasibility over performance parity. The humanoid first-assistant study demonstrated successful completion of a cadaveric sphenoidectomy with stable visualization, but reported no numerical metrics for image stability, tip error, tremor amplitude, contact force, path smoothness, latency, or collisions, and included no baseline comparison against a human assistant or dedicated camera-holder robot (Cho et al., 27 Feb 2026). The voice-control interface reported high recognition performance qualitatively, but did not provide exact recognition percentages in the manuscript text and did not measure workload reduction despite that being a primary motivation (Davila et al., 2024). The gaze-based camera-assistance study lacked significance testing and did not isolate gaze-only contributions (Abdelaal et al., 2023). The needle pick-up system relied on fiducial markers, manual initialization, a fixed endoscope, and free-space needle presentation rather than tissue interaction (D'Ettorre et al., 2018).

This evidentiary pattern suggests that ASAR research currently supports possibility and bounded utility more strongly than clinical superiority or robustness. The field has many proof-of-concept systems that work under structured assumptions, but fewer demonstrations of generalization across anatomies, surgeons, procedures, and operating-room contingencies.

6. Open problems, misconceptions, and development trajectory

A recurrent misconception is that surgical autonomy should be measured primarily by how close a system is to an autonomous surgeon. The ASAR literature argues instead for a staged path centered on assistant roles. One roadmap decomposes progress into instruction, execution, planning and coordination, navigation, emergency handling, self-doubt, observer, and innovator “engines,” implying that clinically useful autonomy will emerge from capability accumulation rather than a single jump to full procedure autonomy (Singh, 2022). The broader autonomy reviews are equally explicit that full-autonomy surgical robots do not currently exist in the literature and that supervised autonomy is the most realistic near-term regime (Thai et al., 2020, Yip et al., 2017).

Another misconception is that embodiment or form factor alone solves the assistant problem. The humanoid endoscope study reports that the robot occupied the “isocenter” of a narrow surgical corridor, whereas a skilled human assistant would typically keep the endoscope more peripheral to preserve surgeon dexterity (Cho et al., 27 Feb 2026). This shows that humanlike morphology does not automatically imply humanlike task geometry. A plausible implication is that ASAR embodiment must be judged not only by reachability, but by whether the robot can stay out of the way while maintaining visibility.

The open problems identified across the literature are consistent. Preference alignment remains difficult because surgeons differ in viewing preferences, manipulation style, and timing. Procedural awareness remains limited because most systems still rely on narrow task assumptions or predefined phase models. Human–robot information exchange remains fragile, especially in noisy or cluttered operating rooms. Skill acquisition in shared workspaces remains data-hungry and poorly standardized (Colan et al., 15 Jul 2025). Safety-specific requirements repeatedly mentioned include elimination of erratic motion, pre-failure warnings, graceful abort behavior, force regulation, active constraints, and transparent communication of failure states (Cho et al., 27 Feb 2026).

The staged development path proposed by the most recent embodiment work is therefore characteristic of the field: first establish that the robot can physically occupy the assistant role; then improve mechanics, force awareness, and workflow compatibility; then automate constrained subtasks with high structure and clear reward signals, especially diagnostic scoping and routine endoscope repositioning; and only afterward move toward context-aware, human-like assistant behavior that reduces verbal command load and can be supervised safely in the operating room (Cho et al., 27 Feb 2026). In that sense, ASARs are less a single technology than a convergent research program in which perception, control, interaction design, and clinical workflow modeling are progressively assembled into physically embodied teammates.