GASv2: Visuomotor Learning for Surgical Grasping

• GASv2 is a visuomotor learning framework for surgical grasping that employs a world-model-based architecture and hybrid control to ensure safety and precision in minimally invasive settings.
• It uses a single stereo endoscopic camera and sim-to-real training with domain randomization to overcome visual ambiguities and achieve a 65% success rate in both phantom and ex vivo scenarios.
• The system generalizes to unseen surgical objects and gripper types without retraining, addressing the limitations of traditional pose estimation and handcrafted visual features.

Grasp Anything for Surgery V2 (GASv2) is a visuomotor learning framework for surgical grasping introduced in "Visuomotor Grasping with World Models for Surgical Robots" (Lin et al., 15 Aug 2025). It targets robot-assisted surgery (RAS), where grasping is a fundamental task and where automation is intended to reduce surgeon workload while enhancing efficiency, safety, and consistency beyond teleoperated systems. GASv2 is defined by three stated objectives: sim-to-real transfer of visuomotor policies to ex vivo surgical scenes, visuomotor learning using only a single stereo camera pair as the standard RAS setup, and object-agnostic grasping with a single policy that generalizes to diverse, unseen surgical objects without retraining or task-specific models. The reported system combines a world-model-based architecture, a surgical perception pipeline for visual observations, and a hybrid control system for safe execution, and it is reported to achieve a 65% success rate in both phantom-based and ex vivo settings (Lin et al., 15 Aug 2025).

1. Definition and problem setting

Published on 2025-08-15, GASv2 is positioned within surgical robotics rather than general robotic grasping, and its stated scope is correspondingly narrow and technically demanding (Lin et al., 15 Aug 2025). The motivating claim is that most prior approaches in robot-assisted surgery rely on explicit object pose tracking or handcrafted visual features. In the GASv2 description, those design choices are said to limit generalization to novel objects, reduce robustness to visual disturbances, and weaken the ability to handle deformable objects.

The system is framed around the specific perceptual and control difficulties of RAS. The available description identifies low signal-to-noise ratio in visual observations, demands for high safety and millimeter-level precision, and the complex surgical environment as the central obstacles to deploying visuomotor learning in practice. These constraints matter because surgical grasping is not simply a grasp-detection problem; it is an embodied closed-loop task performed under severe sensing ambiguity, narrow workspace tolerances, and safety-critical execution requirements.

Within that setting, GASv2 emphasizes object-agnostic operation. The relevant claim is not merely that it can grasp several objects, but that a single policy is intended to generalize to diverse, unseen surgical objects without retraining or task-specific models. This makes the framework notable as a surgical manipulation system aimed at policy unification rather than object-wise specialization (Lin et al., 15 Aug 2025).

2. Core technical characterization

The explicit technical description of GASv2 consists of three components: a world-model-based architecture, a surgical perception pipeline for visual observations, and a hybrid control system for safe execution (Lin et al., 15 Aug 2025). The first component places GASv2 in the line of visuomotor systems that do not reduce control to explicit pose-estimation pipelines. The second indicates that perception is tailored to surgical imagery rather than generic tabletop vision. The third makes safety a first-class systems property rather than a downstream evaluation criterion.

The sensor configuration is also central to the identity of GASv2. The framework is designed to operate using only a single pair of endoscopic cameras, described as the standard RAS setup. This is an important constraint because many high-performing robotic grasping pipelines depend on additional sensing, multiple views, or external object-state estimation. GASv2 instead asserts that surgical visuomotor grasping can be learned and deployed under the canonical stereo endoscopy regime (Lin et al., 15 Aug 2025).

The available description does not specify the latent-state representation, transition model, training losses, controller decomposition, or policy parameterization. Accordingly, the technical meaning of the world model in GASv2 is stated only at the architectural level. What is clear is the systems-level arrangement: visual observations are processed through a surgical perception pipeline, policy learning is organized around a world-model-based architecture, and action execution is mediated by a hybrid control layer intended to preserve safe behavior.

3. Training regime and sim-to-real transfer

GASv2 is trained in simulation using domain randomization for sim-to-real transfer, then deployed on a real robot in both phantom-based and ex vivo surgical settings (Lin et al., 15 Aug 2025). This training route is part of the problem definition rather than an implementation detail: one of the three key challenges explicitly named by the paper is sim-to-real transfer of visuomotor policies to ex vivo surgical scenes.

The use of domain randomization indicates that transfer robustness is treated as a learned property, not merely as an evaluation add-on. In the GASv2 formulation, the target domain is not restricted to controlled benchtop phantoms but includes ex vivo scenes, which are materially closer to surgical handling conditions than purely synthetic or rigid tabletop settings. The reported deployment on a real robot further situates GASv2 as a sim-trained, real-executed manipulation system rather than a simulation-only benchmark (Lin et al., 15 Aug 2025).

A second training-time implication follows from the stated sensor restriction. Because GASv2 uses only a single stereo camera pair, sim-to-real transfer must preserve competence under the same limited visual observability available in standard RAS. This suggests that the framework is organized around practical endoscopic sensing assumptions rather than around richer but less clinically typical instrumentation. A plausible implication is that the sim-to-real burden in GASv2 is inseparable from the monocular-stereo ambiguity, occlusion, and scene complexity of minimally invasive surgery.

4. Empirical behavior, generalization, and robustness

The principal empirical claim is that extensive experiments show a 65% success rate in both phantom-based and ex vivo surgical settings (Lin et al., 15 Aug 2025). The available description treats this as a joint demonstration of performance across two distinct deployment conditions rather than as an isolated phantom result.

The same description reports three additional behavioral properties. First, GASv2 generalizes to unseen objects. Second, it generalizes to unseen grippers. Third, it adapts to diverse disturbances. Taken together, these claims define the system’s intended contribution more precisely than the headline success rate alone: the point is not only that the policy works, but that it does so under changes in object identity, end-effector configuration, and perturbation regime without retraining or task-specific model substitution (Lin et al., 15 Aug 2025).

Reported property	GASv2 statement
Deployment settings	Phantom-based and ex vivo surgical settings
Success rate	65% in both settings
Object generalization	Generalizes to unseen objects
End-effector generalization	Generalizes to unseen grippers
Disturbance handling	Adapts to diverse disturbances

These results position GASv2 as a robustness-oriented surgical grasping system. In surgical manipulation, robustness is not a secondary desideratum, because scene changes, visual disturbances, and instrument variation are routine. The reported ability to maintain one policy across those factors is therefore central to the framework’s identity.

5. Relation to GAS and adjacent research

The designation “V2” suggests continuity with the earlier GAS framework introduced in "World Models for General Surgical Grasping" (Lin et al., 2024). That earlier system proposed a world-model-based deep reinforcement learning framework for general surgical grasping, used a DreamerV2 backbone, introduced uncertainty-aware depth estimation for inaccurate depth regions, encoded depth and mask information into a compact $64 \times 64 \times 3$ Dynamic Spotlight Adaptation representation, and reported an average success rate of 69% across six object categories on a real robot (Lin et al., 2024). GASv2, by contrast, is explicitly described through its single stereo endoscopic setup, ex vivo deployment, hybrid control for safe execution, and object-agnostic single-policy formulation (Lin et al., 15 Aug 2025).

Aspect	GAS	GASv2
Core description	World-model-based deep reinforcement learning for general surgical grasping	Visuomotor learning framework for surgical grasping
Visual setup	RGB-D with compact DSA representation	Single pair of endoscopic cameras
Transfer setting	Sim-to-real on a real robot	Sim-to-real to phantom-based and ex vivo settings
Reported real-robot headline	Average success rate: 69%	65% success rate in both settings

This comparison does not by itself establish a complete architectural delta between the two systems, because the GASv2 description available here does not enumerate internal modules at the same level of detail as GAS. What it does establish is a shift in emphasis. GASv2 foregrounds ex vivo deployment, single-stereo sensing, unseen-gripper generalization, and disturbance adaptation as its defining claims.

A broader neighboring context comes from deformable-object grasping research. "DefGraspSim: Physics-based simulation of grasp outcomes for 3D deformable objects" argues that grasping deformable objects requires attention to deformation, stress, strain energy, and stability rather than only contact kinematics (Huang et al., 2022). GASv2’s abstract does not report such field-level metrics, but its emphasis on ex vivo scenes and on the limitations of prior methods for deformable objects places it within the same larger research problem: safe and robust manipulation of non-rigid surgical targets.

6. Interpretation, limitations, and open questions

Several aspects of GASv2 are clearly established. It is a world-model-based surgical visuomotor framework; it is trained in simulation with domain randomization; it operates from a single stereo endoscopic pair; it is deployed on a real robot; and it is reported to achieve object-agnostic grasping with a 65% success rate in both phantom-based and ex vivo settings while generalizing to unseen objects and grippers and adapting to diverse disturbances (Lin et al., 15 Aug 2025).

At the same time, the currently available description does not specify detailed architecture internals, objective functions, simulator settings, action parameterization, controller frequencies, ablation structure, or failure taxonomy. As a result, firm statements can be made about system scope, sensing assumptions, transfer setting, and headline outcomes, but not about the exact computational mechanisms by which the world model, perception stack, and hybrid controller interact.

This bounded visibility also matters for interpretation. GASv2 should not be misconstrued as a pose-estimation-driven or handcrafted-feature system; its stated motivation is explicitly to move beyond those approaches. Nor should it be read as a multi-view or task-specific architecture; the stated setup uses only a single pair of endoscopic cameras and a single policy without retraining or task-specific models. A plausible implication is that GASv2 is best understood as an attempt to consolidate surgical grasping into a unified visuomotor control framework under realistic RAS sensing constraints, with safety handled through hybrid control rather than through perception alone (Lin et al., 15 Aug 2025).