← Back to Research

Autonomous Ice Resurfacing

Dissertation Research · In Progress

Isaac SimPhysXGymnasiumROS 2NSGA-IIPythonReinforcement Learning

Project media — Isaac Sim screenshot or demo video

Problem

Ice resurfacing (Zamboni operation) is a repetitive, precision-critical task performed multiple times daily in ice rinks worldwide. Current operations rely entirely on human operators following fixed patterns, with no adaptation to actual ice conditions. There is no public simulation benchmark for testing autonomous resurfacing under realistic dynamic ice degradation, and existing work focuses on either control (making the vehicle drive) or geometric coverage without integrating environmental awareness.

Approach

This research develops a complete autonomous ice resurfacing pipeline across three integrated components:

  • Isaac Sim Rink Generator: A procedural pipeline for realistic ice rink simulation including PhysX heightfield topology with skating ruts and wear zones, PBD snow particle systems, and a sensor suite (RGBD camera + RTX LiDAR) with material-specific responses.
  • Gymnasium RL Environment: A 2D simulation platform with a regulation NHL rink (60.96m x 25.9m), dynamic bicycle model with realistic ice friction (~0.05 mu), pluggable dynamics architecture, and a deterministic coverage path baseline policy.
  • Uncertainty-Aware Adaptive Planner: A hybrid planner coupling resurfacing demand (rut depth, snow layer, refreeze readiness) with localization confidence. Uses multi-resolution demand mapping with bounded online local updates and NSGA-II for multi-objective trajectory optimization.

Research Questions

  1. How does fixed coverage perform under degraded perception and slip-prone conditions?
  2. Does adaptive coverage to ice-quality demand improve resurfacing quality without efficiency loss?
  3. Does localization-confidence awareness improve robustness vs quality-only adaptation?
  4. What tradeoff emerges between efficiency and robustness under degradation?

Current Status

  • Isaac Sim rink generator pipeline complete (phases A1, A2, B)
  • Gymnasium environment with 171+ passing tests across 6 test layers
  • Deterministic Boustrophedon coverage baseline implemented
  • NSGA-II multi-objective optimization framework in development
  • Targeting ICRA/IROS for full paper submission