Surgical Steering: Systems & Applications

• Surgical steering is the deliberate control of instruments or model behaviors to follow curved, anatomically optimized trajectories beyond rigid straight-line paths.
• It integrates advanced sensing, calibration, and control methods with tools like steerable needles and curved drills to navigate complex, obstacle-rich environments.
• Recent approaches extend its application to machine learning, where targeted interventions modulate internal representations with precise, measurable improvements in performance.

Surgical steering denotes the deliberate control of a tool, conduit, or energy-delivery path so that it follows a clinically or mechanically preferred trajectory rather than the straight-line path imposed by conventional rigid instruments. In interventional medicine, the term covers steerable needles, concentric-tube drills, microcatheters, endoscopes, balloons, and laser-guidance systems that are designed to navigate obstacle-rich or anatomically constrained spaces while preserving safety margins and targeting accuracy (Kuntz et al., 2022, Kulkarni et al., 2 Jul 2025). In more recent mechanistic machine-learning literature, the same phrase is used analogously for sparse, targeted interventions on internal model representations or policy actions, where the object being steered is a behavior expressed by the model rather than a physical instrument (Cristofano, 13 Jan 2026, Sankaranarayanan et al., 17 Feb 2026).

1. Definition, scope, and recurring design pattern

Across the medical robotics literature, surgical steering is defined by intentional trajectory control under constraints imposed by anatomy, tissue interaction, imaging limitations, and safety. In orthopedic fixation, it is described as deliberately controlling the trajectory of a cutting instrument within bone to follow an anatomically optimal path rather than the straight-line path imposed by rigid tools (Kulkarni et al., 2 Jul 2025). In lung biopsy, the same idea appears as autonomous intra-tissue needle navigation to intraparenchymal targets while avoiding bronchi, clinically significant blood vessels, and the visceral pleura (Kuntz et al., 2022). In middle-ear laser surgery, it appears as automatic steering of a laser spot to achieve exhaustive residual ablation inside a millimetric cavity (So et al., 2022).

A recurring systems pattern is visible across otherwise different embodiments. Each system couples a steering mechanism, a geometric or biomechanical model, a sensing stack, a registration or calibration procedure, and a control policy that translates desired task-space motion into actuator commands. This architecture is explicit in autonomous needle steering with fiducial registration and airway ICP (Kuntz et al., 2022), in concentric-tube drilling frameworks with pivot and hand–eye calibration (Sharma et al., 2024, Maroufi et al., 2 Jul 2025), and in image-based endoscopic orientation control using a calibrated pixel-ratio surrogate for balloon bending angle (McCandless et al., 3 Nov 2025).

Domain	Representative systems	Steering variable
Soft-tissue access	Flexure-tip needle, PBN, OncoReach stylet (Kuntz et al., 2022, Aktas et al., 2023, Kheradmand et al., 20 Jan 2026)	Curvature, bevel orientation, tendon-induced bending
Hard-tissue and energy delivery	CT-SDR drilling, MRI-driven neuroendoscope, steerable laser systems (Kulkarni et al., 2 Jul 2025, III et al., 2022, So et al., 2022)	Drill path, tip orientation, laser spot trajectory
Mechanistic model intervention	SRA, GCM, per-vulnerability code steering (Cristofano, 13 Jan 2026, Sankaranarayanan et al., 17 Feb 2026, Sandoval et al., 17 Apr 2026)	Internal representation subspace or action direction

This breadth shows that surgical steering is not a single mechanism but a systems concept. A plausible implication is that the term now denotes a design objective—localized, controllable deviation from a default straight or unstructured behavior—more than any particular hardware or algorithmic substrate.

2. Soft-tissue steering modalities

Needle steering in soft tissue is the most explicit classical form of surgical steering. In "Autonomous Medical Needle Steering In Vivo," a laser-patterned, highly flexible bevel-tip needle is used to navigate through lung parenchyma along curvilinear trajectories, with curvature direction controlled by axial rotation and execution segmented into 10 mm insertion arcs synchronized to ventilator breath holds at peak tidal volume (Kuntz et al., 2022). The same paper defines surgical steering broadly enough to include steerable catheters, continuum robots, and flexible needles, but its distinctive result is the first autonomous in vivo obstacle-avoiding needle navigation to intra-tissue targets in a living organ without visual line-of-sight inside parenchyma (Kuntz et al., 2022).

A simpler but formally structured abstraction appears in timed-game synthesis for bevel-tip needles in gelatin phantoms. There, the needle is modeled as a non-holonomic system with constant-curvature motion and discrete rotate/no-rotate decisions:

$$\dot{x} = v \cos\theta,\quad \dot{y} = v \sin\theta,\quad \dot{\theta} = \sigma\,\kappa\,v.$$

The model is deliberately geometric rather than force-based, and a preprocessing stage estimates the actual initial tip angle and position before strategy synthesis; measured-trace experiments reported a maximum deviation of 1.84 mm (Rogalla et al., 2020). This line of work treats steering as a formally synthesizable reachability problem rather than a purely heuristic manipulation task.

Several devices realize curvature through programmable local asymmetry rather than bevel mechanics. The programmable-bevel tip needle (PBN) uses four interlocking segments whose relative axial offsets create a composite asymmetric tip; synchronous insertion of the offset configuration produces controlled curvature, and experimental characterization compared a 2.5 mm extrusion-manufactured device with a 1.3 mm thermally drawn device (Aktas et al., 2023). The TD-PBN preserved complex multi-lumen geometry at smaller outer diameter, but the stiffer polycarbonate material reduced steering performance: maximum single-leading-segment curvature was $0.0242 \pm 0.003\ \mathrm{mm}^{-1}$ for the 2.5 mm EM-PBN and $0.0092 \pm 0.001\ \mathrm{mm}^{-1}$ for the 1.3 mm TD-PBN (Aktas et al., 2023).

Soft-tissue steering also includes tendon-driven stylets and active base reorientation. The OncoReach stylet for cervical brachytherapy uses a nitinol backbone, spherical joints, and routing disks inside standard 15- and 13-gauge needles; free-space tip deflection increased with lower gauge stiffness, higher joint count, and larger tendon moment arm, and the clinically chosen configuration was a 15-gauge needle with a 30-joint tip section and symmetric routing disks (Kheradmand et al., 20 Jan 2026). In a patient-derived multi-composite uterus/pelvis phantom, the expert user steered from a more central entry to the lateral-most mid-plane target, with measured lateral tip deflections of 12.6 mm anteriorly, 15.8 mm medially, and 24.0 mm posteriorly (Kheradmand et al., 20 Jan 2026).

By contrast, the MRI-visible breast-biopsy system relies on a hand-mounted motorized angulation frame rather than an intrinsically curving needle. A 2-DOF differential bevel-gear stage steers a rigid needle by azimuth and elevation updates derived from the vector between the current EM-tracked tip and a lesion re-localized through stereo marker tracking and thin-plate-spline deformation mapping (Lagomarsino et al., 2021). This is still surgical steering in the task sense: the path is adjusted online to converge to a moving lesion under tissue deformation even though curvature is generated by piecewise-linear re-aiming rather than by continuum mechanics.

3. Hard-tissue steering, curved drilling, and steerable energy delivery

In bone, surgical steering primarily appears as curved drilling. The concentric tube steerable drilling robot (CT-SDR) uses precurved superelastic NiTi tubes, a flexible torque-transmission chain, and independent translation/rotation actuation to create non-linear tunnels that rigid drills cannot realize (Kulkarni et al., 2 Jul 2025, Sharma et al., 2023). In spinal fixation, a biomechanics-aware framework combined patient-specific QCT-derived finite-element analysis with CT-SDR design, and the selected curved trajectory reduced maximum stress and strain in cancellous bone by approximately 80.1% and 77.5%, respectively, relative to a straight trajectory (Sharma et al., 2023). The same paper reported curvature reproduction errors between 1.7% and 2.2% in core tests on Sawbone PCF 10 blocks (Sharma et al., 2023).

The hard-tissue steering literature has expanded from planar J-curves to J-, U-, S-, and out-of-plane tunnels. The spinal-tumor CT-SDR demonstrated smooth planar and out-of-plane J-shape branch drilling, U-shape drilling, and cavity drilling in simulated bone; the U-shape path reached a point 82 mm perpendicular to the entry trajectory with an orientation change of 153°, and pure rotational cavity drilling in 10 PCF sawbone increased cavity diameter from 10.34 mm to 26.64 mm over a 30.03 mm insertion distance (Sharma et al., 2023). In pelvic fixation, the 4-DoF pelvic CT-SDR created long, smooth, planar and out-of-plane S-shaped drilling trajectories in bone phantoms under fluoroscopy, with total tunnel length approximately 90 mm and average drilled diameter approximately 7.1 mm (Kulkarni et al., 2 Jul 2025).

Autonomous robotic integration adds another layer. A 7-DoF KUKA LBR Med integrated with CT-SDR was used for autonomous J-shaped drilling in osteoporotic vertebral phantoms, with a planned curvature radius of 69.5 mm and measured curvature errors of 1.03%, 1.29%, and 4.68% across mounting conditions (Sharma et al., 2024). The later S3D framework generalized this to a four-phase calibration, registration, and navigation workflow for planar and out-of-plane steerable drilling, reporting a measured radius of curvature of $70.83 \pm 1.72\ \mathrm{mm}$, or 1.9% error versus nominal 69.5 mm (Maroufi et al., 2 Jul 2025).

Surgical steering in hard or constrained spaces also includes steerable energy delivery. In middle-ear surgery, a macro–micro robotic platform automatically generated intra-region sinusoidal laser-scanning paths and inter-region linking paths, then tracked them with image-based control; this achieved a mean spot position error of 85.78 $\mu$m and 98.12% average area coverage of residual regions (So et al., 2022). At a smaller distal scale, the tendon-actuated galvanometer (TAG) combined continuum access with free-beam laser steering, reaching $30.14 \pm 0.90^\circ$ mirror rotation, equivalent to approximately $60.28^\circ$ laser reflection angle, with a forward-kinematics endpoint RMSE of 0.68 mm at 8.56 mm standoff (Yamamoto et al., 2024).

4. Sensing, registration, modeling, and closed-loop control

Surgical steering systems depend on a sensing and registration stack that links a desired target or path to the tool state under deformation, motion, or calibration uncertainty. In autonomous lung biopsy, preoperative CT acquired under breath hold was combined with fiducial registration and a correspondence-free airway ICP refinement, while bronchoscope and steerable needle tips were continuously tracked by electromagnetic coils and overlaid on CT anatomy (Kuntz et al., 2022). Respiratory motion was handled by safe insertion time windows that started at peak tidal volume and limited measured displacements to under 1 mm during short breath holds (Kuntz et al., 2022).

In breast biopsy, real-time lesion localization was maintained by mapping MRI-visible and optically tracked surface markers through thin-plate splines. The deformable map $u(x)$ follows the standard TPS structure $u(x)=Ax+b+\sum_i w_i\phi(\|x-p_i\|)$, and the lesion estimate in tool coordinates was updated continuously before the motorized angulation frame re-aimed the needle (Lagomarsino et al., 2021). This workflow explicitly separates rigid pose recovery for the hand-mounted device from nonrigid target recovery for the breast.

Concentric-tube and drilling systems rely heavily on calibration chains. Both the autonomous spinal fixation framework and S3D used pivot calibration to recover tool-tip offsets and single-step robot-world/hand–eye calibration of the $AX = ZB$ form to relate robot, tracker, and tool frames (Sharma et al., 2024, Maroufi et al., 2 Jul 2025). In S3D, desired drill-tip poses were marked with an optical digitizer and converted into robot end-effector goals through chained transforms, after which the rigid pilot and flexible J-shaped phases were executed sequentially (Maroufi et al., 2 Jul 2025).

Control laws vary by actuation modality. The steerable balloon endoscope used image-derived pixel ratio

$0.0242 \pm 0.003\ \mathrm{mm}^{-1}$0

as a surrogate for bending angle, then drove a multi-threshold bang-bang controller on syringe-pump motor speed to regulate orientation during tool insertion and removal (McCandless et al., 3 Nov 2025). The MRI-driven neuroendoscope solved a real-time inverse-kinematics Cosserat rod model via Levenberg–Marquardt, with steering torque generated by Lorentz-force microcoils under the MRI static field:

$0.0242 \pm 0.003\ \mathrm{mm}^{-1}$1

Experimentally identified flexural rigidity $0.0242 \pm 0.003\ \mathrm{mm}^{-1}$2 grounded the controller’s predictions up to 90° bends (III et al., 2022).

The hard-tissue concentric-tube literature is grounded in Cosserat rod mechanics even when full online solvers are not deployed. The pelvic CT-SDR paper states the classical kinematic and equilibrium relations

$0.0242 \pm 0.003\ \mathrm{mm}^{-1}$3

$0.0242 \pm 0.003\ \mathrm{mm}^{-1}$4

with constitutive law

$0.0242 \pm 0.003\ \mathrm{mm}^{-1}$5

and energy

$0.0242 \pm 0.003\ \mathrm{mm}^{-1}$6

These equations frame steering as curvature synthesis under stiffness, precurvature, and interaction constraints (Kulkarni et al., 2 Jul 2025).

5. Validation, safety, and workflow implications

The most mature validations combine targeting metrics with explicit safety outcomes. In vivo porcine lung biopsy achieved final targeting errors of 3.4 mm, 1.8 mm, and 2.7 mm over trajectory lengths of 43 mm, 48 mm, and 38 mm, respectively, with no pneumothorax, atelectasis, vessel perforation, or hemorrhage confirmed by post-deployment CT (Kuntz et al., 2022). In the same study’s ex vivo comparison with manual clinical bronchoscopy, the autonomous steerable-needle system achieved insertion length $0.0242 \pm 0.003\ \mathrm{mm}^{-1}$7 mm versus $0.0242 \pm 0.003\ \mathrm{mm}^{-1}$8 mm and targeting error $0.0242 \pm 0.003\ \mathrm{mm}^{-1}$9 mm versus $0.0092 \pm 0.001\ \mathrm{mm}^{-1}$0 mm, with reported $0.0092 \pm 0.001\ \mathrm{mm}^{-1}$1 and $0.0092 \pm 0.001\ \mathrm{mm}^{-1}$2, respectively (Kuntz et al., 2022).

Accuracy improvements are also reported in deformation-aware percutaneous workflows. The MRI-guided breast-biopsy platform achieved mean lesion localization error 1.16 mm with TPS versus 1.34 mm with rigid-only mapping, and end-to-end targeting accuracy of 2.21 mm mean 3D distance between needle tip and lesion (Lagomarsino et al., 2021). This is clinically notable because lesion displacement between preoperative and intraoperative states averaged 4.30 mm, so the steering system operated in a regime where deformation compensation, not nominal image-to-image correspondence, determined feasibility (Lagomarsino et al., 2021).

In endoscopic orientation control, performance is often expressed as stabilization under disturbance. The steerable balloon endoscope produced deployed optical face diameters of 8–11 mm before bending occurred, then achieved bending up to approximately 100° as inflation volume increased; closed-loop control reached 60° in under 6 s with average tip velocity approximately $0.0092 \pm 0.001\ \mathrm{mm}^{-1}$3s and maintained orientation within $0.0092 \pm 0.001\ \mathrm{mm}^{-1}$4 during tool insertion and removal (McCandless et al., 3 Nov 2025). The MRI-driven neuroendoscope demonstrated up to 100° realized tip orientations, grasping forces up to 31 mN, and steering at total power approximately 0.4663 W, while the paper explicitly noted a discrepancy in the reported thermal-ablation power figures (III et al., 2022).

Hard-tissue steering studies emphasize breach avoidance, curvature fidelity, and geometric compatibility with implants. The pelvic CT-SDR reported outer and inner arc-length errors of 2.7% and 6.2%, inner curvature radius $0.0092 \pm 0.001\ \mathrm{mm}^{-1}$5 mm versus ideal 50 mm, and average insertion drilling time of 55 seconds for 90 mm S-shaped tunnels (Kulkarni et al., 2 Jul 2025). S3D reported tip position error $0.0092 \pm 0.001\ \mathrm{mm}^{-1}$6 mm for the rigid tip and $0.0092 \pm 0.001\ \mathrm{mm}^{-1}$7 mm for the flexible tip, orientation errors below 1° on all axes, and no cortical breach observed on X-ray in the reported trials (Maroufi et al., 2 Jul 2025).

A recurrent misconception is that steering implies full autonomy. The literature does not support that simplification. The lung-biopsy system is explicitly semi-autonomous, combining manual bronchoscope navigation, teleoperated aiming, and fully autonomous needle steering (Kuntz et al., 2022). The breast-biopsy tool preserves one-handed radiologist control and haptic feedback while only enabling computer guidance when a momentary button is pressed (Lagomarsino et al., 2021). The OncoReach stylet similarly uses a reusable handheld manual interface, with wrist rotation setting steering direction and finger pull providing actuation (Kheradmand et al., 20 Jan 2026). Steering, in other words, is compatible with supervisory autonomy, teleoperation, or shared control.

6. Terminological expansion to mechanistic steering in machine learning

Recent arXiv papers extend the term surgical steering from physical intervention to internal model intervention. In "Surgical Refusal Ablation," surgical steering is defined as the precise suppression or amplification of a behavior by acting on a low-dimensional representation subspace while keeping the rest of the model’s semantic geometry and competencies intact and avoiding distribution drift (Cristofano, 13 Jan 2026). The method computes a raw refusal direction

$0.0092 \pm 0.001\ \mathrm{mm}^{-1}$8

then residualizes it against a matrix $0.0092 \pm 0.001\ \mathrm{mm}^{-1}$9 of Concept Atoms:

$70.83 \pm 1.72\ \mathrm{mm}$0

and applies either activation steering

$70.83 \pm 1.72\ \mathrm{mm}$1

or a rank-one weight update

$70.83 \pm 1.72\ \mathrm{mm}$2

Across five models, the paper reported refusal reduction to 0–2%, mean $70.83 \pm 1.72\ \mathrm{mm}$3PPL on wikitext_2_raw of approximately +0.02, and mean first-token KL approximately 0.025 (Cristofano, 13 Jan 2026).

"Surgical Activation Steering via Generative Causal Mediation" uses the phrase for sparse interventions on attention heads selected by their mediated effect on long-form outputs (Sankaranarayanan et al., 17 Feb 2026). For a head $70.83 \pm 1.72\ \mathrm{mm}$4, the mediator score is

$70.83 \pm 1.72\ \mathrm{mm}$5

with indirect effect estimated by activation patching between contrastive prompt–response pairs. Steering then modifies selected head outputs through difference-in-means, mean steering, or ReFT. The paper reported that activation patching and attribution patching each reached 0.40 average success across models and tasks, significantly outperforming ITI probes and random baselines under FDR-adjusted $70.83 \pm 1.72\ \mathrm{mm}$6 (Sankaranarayanan et al., 17 Feb 2026).

A narrower but strongly mechanistic use appears in insecure-code repair. "Surgical Repair of Insecure Code Generation in LLMs" localized the failure to a single final layer and applied per-vulnerability steering vectors at that layer:

$70.83 \pm 1.72\ \mathrm{mm}$7

For Llama-3.1-8B-Instruct, secure generation for CWE-787 improved from 6.7% to 73.3% at $70.83 \pm 1.72\ \mathrm{mm}$8, described as a 74% relative reduction in insecure generation, with steering overhead near zero under persistent application methods (Sandoval et al., 17 Apr 2026).

The term also appears from the opposite direction in adversarial machine vision for surgical robotics. "Adversarial Attacks on Learned Policies for Surgical Robotic Tasks" studies steering attacks in which imperceptible visual perturbations drive end-to-end policies toward attacker-specified action directions, including targeted action optimization

$70.83 \pm 1.72\ \mathrm{mm}$9

Over 560 physical experiments on debridement and suturing, the paper reported an average 61% reduction in surgical subtask success rates, showing that learned policies can themselves be externally steered in dangerous directions (Jin et al., 10 Jun 2026).

This usage is distinct from operative instrument steering because the steered object is an internal representation or learned control policy rather than a physical instrument. At the same time, a plausible implication is that both literatures now use surgical steering to denote the same engineering ideal: targeted intervention at the point of highest leverage, with explicit concern for collateral effects, calibration, and safety.