Reproduce “Learning Smooth State-Dependent Traversability from Dense Point Clouds” on OpenResearch

About this paper

Zihao Dong, Alan Papalia, Leonard Jung, Alenna Spiro, Philip R. Osteen, Christa S. Robison, Michael Everett

Researchers at Northeastern University and DEVCOM Army Research Laboratory introduced SPARTA, a method that learns state-dependent terrain traversability from dense point clouds by modeling angle-dependent risk as a smooth analytical function. This approach achieved a 91% success rate in challenging simulated boulder fields and enabled safer robot navigation on hardware by improving generalization and computational efficiency.

Key insights

• Imposing geometric structure by representing angle-dependent risk with Fourier basis functions significantly improves model generalization to unseen angles and naturally handles angular periodicity.
• Predicting an analytical function's coefficients (rather than direct angle predictions) allows for an order-of-magnitude faster risk query during planning, enhancing computational efficiency for real-time applications.
• Modeling risk probabilistically via categorical distributions and integrating CVaR into the planning objective provides a robust framework for safety-critical tasks by accounting for aleatoric uncertainty and worst-case scenarios.
• direct imposition of dynamics constraints through a novel output parameterization ensures that all generated trajectories are inherently feasible for the robot.
• use of a denoising diffusion process enables efficient generation of diverse, high-quality trajectories that respect complex robot dynamics, outperforming traditional sampling methods.

Summary by alphaXiv.