SpatialWorld Benchmark
Anonymous Repository for NeurIPS 2026 Submission
This is an anonymous dataset repository for peer review purposes. All identifying information has been removed to maintain double-blind review compliance.
Overview
SpatialWorld is a comprehensive benchmark for evaluating embodied AI agents in complex 3D spatial reasoning and manipulation tasks. The benchmark encompasses diverse virtual environments and task types to assess agents' capabilities in understanding and interacting with physical spaces.
Benchmark Components
Environments
The benchmark includes the following simulation environments:
| Environment | Description | Tasks |
|---|---|---|
| AI2-THOR | Indoor household environments | Object manipulation, navigation, and rearrangement tasks |
| CARLA | Autonomous driving simulator | Driving scenarios and vehicle control tasks |
| VirtualHome | Multi-agent household simulator | Multi-step task completion and social interactions |
| ProcTHOR | Procedurally generated homes | Generalization across diverse household layouts |
| EmbodiedCity | Urban environment simulation | City-scale navigation and exploration |
| Game | Puzzle and game environments | Rubik's cube, maze solving, and 3D block puzzles |
Task Structure
Each task in the benchmark contains:
- Task ID: Unique identifier (e.g.,
ai2thor00000) - Instruction: Natural language description of the goal
- Scene/Environment: The specific scene configuration
- Golden Actions: Reference action sequence for task completion
- Success Conditions: Formal criteria for task success evaluation
- Initial State: Starting configuration via
init.json
Data Format
benchmark/
├── ai2thor/tasks/{task_id}/
│ ├── task.json # Task definition and success conditions
│ └── init.json # Initial scene configuration
├── carla/tasks/{task_id}/
│ ├── task.json
│ └── init.json
├── embodiedcity/
│ └── Datasets/
├── game/data/
│ ├── rubik_cube/
│ ├── maze3d/
│ ├── maze3d_pro/
│ └── Block3D/
├── procthor/
│ └── tasks/
└── virtualhome/
└── tasks/