SIPA: Quantifying Physical Integrity and the Sim-to-Real Gap in 7-DoF Trajectories

Researchers have introduced SIPA (Spatial Intelligence Physical Audit), a groundbreaking diagnostic framework designed to quantify physical inconsistencies in AI-generated motion trajectories. The tool operates at the trajectory level, analyzing 7-DoF CSV data without requiring access to source code or internal simulator states, making it compatible with virtually any system that produces spatial motion outputs—from physics simulators like NVIDIA Isaac Sim and MuJoCo to neural world models like OpenAI Sora and robotic foundation models. At its core lies the Non-Associative Residual Hypothesis (NARH), which posits that physical inconsistency stems from discrete solver ordering in parallel constraint resolution rather than just algebraic errors, introducing measurable order-dependent residuals that accumulate as structured drift.

The NARH framework mathematically defines the Non-Associative Residual (R_t) as the norm of an associator measuring path-dependence induced by discrete solver ordering. This residual becomes significant in high-interaction-density regimes like contact-rich robotics, potentially explaining sim-to-real divergence and policy brittleness. SIPA supports three data pathways: Tier 1 (native spatial intelligence from simulators/telemetry), Tier 2 (structured world generators like World Labs Marble), and Tier 3 (experimental video-to-pose pipelines for models like Sora). If validated, NARH suggests that order sensitivity is a structural property of discrete solvers, and associator magnitude could serve as an early-warning indicator for embodied AI training failures in simulation.