Bridging the Reality Gap: Domain Randomization Strategies for Indian Sewer Geometry in Isaac Sim

Abstract

Sim-to-real transfer for robotic navigation in standardized environments benefits from well-characterized geometry and material properties. Indian sewer infrastructure presents a distinct challenge: tunnels are hand-built with irregular cross-sections, walls are constructed from mixed materials (brick, stone, concrete, bare earth), and dimensions vary unpredictably even within a single municipal system. Standard domain randomization approaches that vary texture, lighting, and physics parameters within narrow ranges fail to capture this structural diversity. We present a geometry-aware domain randomization strategy implemented in NVIDIA Isaac Sim that procedurally generates sewer environments matching the statistical distribution of Indian underground infrastructure. Our approach randomizes tunnel cross-section shape (circular, rectangular, trapezoidal, irregular), wall material composition, surface degradation patterns, obstacle placement, and water level. We train SafAI navigation policies in these randomized environments and demonstrate a 41% improvement in zero-shot sim-to-real transfer compared to standard domain randomization when evaluated in three Indian municipal sewer systems. The procedural generation pipeline and all environment assets are released as open-source tools.

Keywords