BioVLN: A Simulation Platform for Visual Language Navigation in Biomedical Laboratories

Most existing embodied navigation platforms, such as Habitat and AI2-THOR, are designed for household or office environments. In these settings, navigation success is typically defined as reaching the center of an object or an arbitrary nearby location. However, biomedical laboratories present unique constraints: instruments like centrifuges or refrigerators have specific operational faces that must be accessed from a particular direction, and dense, workflow-dependent layouts require robots to maintain safe clearance from surrounding equipment. Existing platforms lack the spatial semantics to encode these functional requirements.

BioVLN addresses these limitations by introducing a three-zone operational envelope for every instrument:

1. Physical Body: The space occupied by the instrument itself.
2. Safety Clearance: A 0.25m buffer zone ensuring navigable gaps between equipment.
3. Operational Envelope: A specific region in front of the usable side where the robot must position itself to perform tasks.

This model is integrated into a dual-entry pipeline that supports both procedural scene generation and designer-authored scene imports via the LabScene Annotation Toolkit (LSAT). The platform provides 47 scenes and 1,667 episodes, offering a standardized environment for training and evaluating navigation agents.

BioVLN benchmarks six navigation methods, ranging from geometric frontier exploration to VLM-guided semantic search. The results demonstrate that purely geometric exploration achieves a 74.4–87.5% success rate. By adopting the three-zone model and sampling multiple valid positions within the operational envelope, agents improve their success rate to 83.3–92.5%. Furthermore, this approach significantly reduces unsafe proximity to laboratory equipment, proving that explicit modeling of instrument affordances is critical for laboratory robotics.