Upper gastrointestinal endoscopy often follows a scan-then-revisit workflow, where a clinician must bring a suspicious region back into view after seeing it only briefly. The new EndoVLA-OOV policy treats this as a language-conditioned control problem: it takes anatomical and appearance descriptions and converts them into continuous bending commands for a robotic endoscope, using initial, intermediate, and current images to guide the search.
The authors contribute an open-source vision-language-action dataset with 1,050 episodes across six tasks and about 262,000 observation-action pairs. Physical evaluation covered wet-lens and low-light conditions, two unseen stomach phantoms, and cross-embodiment transfer to an unseen ex-vivo environment. In the cross-embodiment test, the full-data policy achieved 93.3% target acquisition success (28 out of 30 trials).
A controlled training-data comparison showed a sharp benefit from more demonstrations: holding the model, training setup, and test configuration fixed, expanding the corpus from 312 to 1,050 episodes increased success from 40.0% to 90.0%. Instruction and observation interventions further indicated that both language input and temporal context contributed to the evaluated acquisition tasks. The results support language-guided endoscopic revisitation across the tested environments and robot embodiments, while highlighting the value of larger task-specific demonstration sets.