The Living Terrain

Two robotics papers address the same fundamental problem: how does a machine navigate tissue that is alive, moving, and actively responding to the machine’s presence?

Du et al. (arXiv: 2604.01523) demonstrate autonomous control of magnetic millirobots in cardiac flow. An untethered magnetic microbot navigates a heart phantom with pulsatile flow — meaning the fluid pushes against the bot with each heartbeat cycle. Vision-guided control adjusts the external magnetic field in real time to keep the bot on course despite turbulent, periodic perturbation. The target: drug delivery inside a beating heart, where the terrain is a vascular system that is simultaneously the patient’s life support and the delivery obstacle.

Gui et al. (arXiv: 2604.01371) develop AffordTissue, a system that predicts dense affordance maps for surgical tool-tissue interaction during cholecystectomy. Rather than treating tissue as a static obstacle, the system predicts where specific tools can safely contact the tissue and what actions are permissible at each location. The affordance map is action-specific: the same tissue surface has different interaction zones for grasping, cutting, and cauterizing. The tissue isn’t just geometry — it’s a surface of possibilities that depends on what you intend to do.

The structural claim: navigating living systems requires treating the terrain as an active participant, not a passive landscape. The cardiac flow is not a static pipe — it pulses, and the pulsation changes the microbot’s dynamics every fraction of a second. Surgical tissue is not a static surface — it deforms, bleeds, and responds to contact. In both cases, the machine must model the terrain’s own physics, not just its geometry.

This is a departure from classical robotics, where the environment is treated as an obstacle map: walls here, gaps there, move through the free space. In a living body, the obstacle map changes continuously. The walls pulse. The gaps shift. The “free space” has fluid flowing through it. And the map itself responds to the robot’s presence — contact with tissue causes deformation, inflammation, bleeding, or healing, none of which appear in a static model.

Du et al.‘s vision-guided control is particularly interesting because the heart’s pulsatile flow is both the obstacle and the transport mechanism. The bot rides the flow when it’s going the right direction and fights it when it’s not. The control problem isn’t just “get from A to B” — it’s “get from A to B while being pushed by a periodic force that alternately helps and hinders.” This is closer to surfing than navigation.

Gui et al.‘s affordance prediction introduces a different complexity: the same tissue location has different safety profiles depending on what the robot intends to do. Grasping tissue might be safe where cutting would be dangerous. Cauterizing might be appropriate where blunt dissection would cause bleeding. The environment isn’t just “what’s there” — it’s “what can I do here.” This is the affordance framework from ecological psychology, originally developed to describe how animals perceive their environment in terms of action possibilities rather than physical properties.

Both papers implicitly argue that the hard part of medical robotics isn’t miniaturization or precision — it’s understanding. The microbot is small enough. The surgical robot is precise enough. What’s missing is a model of the living environment that captures its dynamics, responsiveness, and action-dependent properties. The machine needs to understand the body, not just fit inside it.

The question this raises: at what point does a surgical or therapeutic robot need to understand biology rather than just geometry and physics? AffordTissue starts answering this — its predictions encode medical knowledge about tissue types, vascular structures, and surgical technique. The cardiac microbot starts answering it differently — its control adapts to hemodynamics without explicitly modeling cardiovascular physiology. Two approaches to the same gap: one explicit (encode the biology), one implicit (learn the physics).