haosulab/SAPIEN

What it solves

SAPIEN is a physics-rich simulated environment designed for robotic vision and interaction tasks. It addresses the need for detailed, part-level understanding of articulated objects, providing a large-scale set of such objects to help robots learn how to interact with complex physical environments.

How it works

SAPIEN provides a simulation engine that integrates physics and rendering. It uses an entity-component system (in version 3.0) and supports PhysX 5 GPU simulation for high-performance physics. The environment includes a specialized renderer (SapienRenderer) that supports ray tracing and GPU-accelerated stereo depth sensor simulation, allowing for realistic visual feedback for robotic agents.

Who it’s for

It is primarily intended for researchers and developers working on robotic vision, embodied AI, and robot learning, particularly those needing to simulate interactions with articulated objects.

Highlights

• Part-level understanding: Built upon the ShapeNet and PartNet datasets to provide detailed articulated objects.
• GPU Acceleration: Supports PhysX 5 GPU simulation and GPU-accelerated stereo depth sensors.
• Advanced Rendering: Features a renderer with ray tracing capabilities and a dedicated render server.
• Flexible Infrastructure: Uses an entity-component system to manage simulation entities and components.