Open Vocab Mobile Manipulation episodes

We describe the fields in our OVMM episodes.

Initial Configuration

rigid_objs: Initial poses of all pickable objects inserted into the scene specified via a 4x4 array
start_position: Position of the agent at the start of the episode.
start_rotation: Orientation of the agent at the start of the episode.

Goal specification

object_category, start_recep_category, and end_recep_category: These fields specify the OVMM task goal - Move object_category from start_recep_category to goal_recep_category

Candidate targets

Each target is specified via an object_id along with information on position and a set of view_points (agent poses from where the target is visible).

candidate_objects: object instances belonging to object_category on top of a receptacle belonging to start_recep_category
candidate_objects_hard: all object instances belonging to object_category
candidate_start_receps: all receptacle instances belonging to start_recep_category
candidate_goal_receps: all receptacle instances belonging to goal_recep_category