New Survey Maps Physics Simulators for Embodied AI Navigation and Manipulation

Physics simulators have become essential tools for training embodied AI systems to navigate environments and manipulate objects without the cost and safety risks of learning directly in the physical world. A newly submitted arXiv survey takes stock of how these simulation platforms address the persistent challenge of sim-to-real transfer — the gap between virtual training and real-world performance.

The Core Problem#

Training robots in physical environments is prohibitively expensive, time-intensive, and often dangerous. Simulation offers an alternative, but robots trained entirely in virtual spaces frequently struggle when deployed in the real world. This fundamental mismatch has spurred researchers to explore what makes certain physics simulators better suited to bridging this divide.

What the Survey Covers#

The review goes beyond earlier analyses by examining simulator properties that have been overlooked in prior work. It catalogs how different platforms support navigation (moving through space) and manipulation (grasping and moving objects), while also documenting their computational demands and hardware requirements. The authors compile a practical resource listing:

• Benchmark datasets for testing and validation
• Performance metrics and evaluation approaches
• Available simulation platforms and their capabilities
• Methods for researchers to identify the right tools for their constraints

Practical Impact#

By providing a structured comparison of physics simulators alongside their resource requirements, the survey aims to help researchers and practitioners make informed choices when selecting simulation environments for embodied AI work.