VIENNA performs robust on unseen environments, through learning task-shared, general representations.

However, the automation of MIS and RAMIS through the use of AI has been slow due to difficulties in data acquisition and curation, partially caused by the ethical considerations of training, testing and deploying AI models in medical environments.

We thank all the reviewers for their helpful feedback and positive view of our work. To address the re-1 viewers concerns (R3,R4,R5), we have added a comparison to Duan et al.'s One Shot Imitation learning in2 Tables 3 and 2, a comparison to a non-normalized TECNet ablation, as well as an evaluation on a Viz-3 Doom navigation task in Table 1. We believe that these additions address all of the main reviewer concerns.4 We evaluate our method in ViZDoom where the goal is to visit waypoints in apredetermined order. Istheembedding14 normalized in the same way in your model?We do not15 normalizetheCPVembeddings. Toteaseaparttheeffects16 of normalization, we have added a comparison to non-17 normalized versions ofTECNet, which are labeled "TE"18 (taskembedding)inTable3and2.19

Similartoaudio-visual navigationtasks,thegoalofourembodied agentistolocalize anaudioeventvia navigating the 3D visual world; however, the agent may also seek help from a human (oracle), where the assistance is provided in free-form natural language.

Explicitmapsalsoallow analytical path planning and thus lead to significantly better exploration and sample complexity.