SCAPE study tests scenario-specific evaluation for robot policies

A new SCAPE research paper describes a framework for estimating robot-policy performance by scenario using limited paired simulation and real-world samples alongside larger simulation rollouts. The method corrects simulation labels for sim-to-real bias and calibrates uncertainty with conformal prediction. In sim-to-sim tests covering autonomous driving and quadruped velocity tracking, the authors report lower scenario-level prediction error than listed baselines. They also evaluated a velocity-tracking policy on a physical Unitree Go2 robot.