Back to datasets
Dataset assetOpen Source CommunityRobot ManipulationAutomated Tasks

ai-habitat/OVMM_episodes

The Open Vocab Mobile Manipulation (OVMM) dataset involves mobile manipulation tasks with initial configurations, goal specifications, and candidate goals. The initial configuration includes the initial poses of objects, the agent's start position and orientation. The goal specification defines moving objects of a specific category from a source container to a target container. The candidate goals section details instances of objects and containers related to the task.

Source
hugging_face
Created
Nov 28, 2025
Updated
Sep 13, 2023
Signals
86 views
Availability
Linked source ready
Overview

Dataset description and usage context

Open Vocab Mobile Manipulation Episodes

Initial Configuration

  • rigid_objs: Initial poses of all pickable objects, specified by 4 × 4 matrices.
  • start_position: Agent’s position at the start of an episode.
  • start_rotation: Agent’s orientation at the start of an episode.

Goal Specification

  • object_category, start_recep_category, and end_recep_category: These fields define the OVMM task goal – moving an object_category from start_recep_category to goal_recep_category.

Candidate Goals

Each goal is specified by an object_id together with a position and a set of view_points (agent poses from which the goal is visible).

  • candidate_objects: Instances of object_category placed on containers belonging to start_recep_category.
  • candidate_objects_hard: All instances of object_category.
  • candidate_start_receps: All container instances belonging to start_recep_category.
  • candidate_goal_receps: All container instances belonging to goal_recep_category.
Need downstream help?

Pair the dataset with AI analysis and content workflows.

Once the source passes your review, move straight into summarization, transformation, report drafting, or presentation generation with the JuheAI toolchain.

Explore AI studio