Sim-Real-Drift-Check-Perturbation-Dataset

150 randomized sensor-noise and physics-perturbation configurations evaluated against a trained CartPole control policy, measuring performance degradation to study the sim-to-real gap in robotics.

About Dataset

This dataset contains 150 randomly sampled perturbation configurations (sensor noise level, motor force scaling, pole length scaling) applied to a linear control policy trained via the cross-entropy method on a from-scratch CartPole simulation. Each row records the resulting performance degradation and a risk category (low/medium/high), generated to study the 'sim-to-real gap' - a well-documented robotics problem where controllers trained in simulation fail when deployed on real hardware due to imperfect sensors and physics mismatches (Tobin et al., 'Domain Randomization for Transferring Deep Neural Networks from Simulation to the Real World', IROS 2017). A key finding from this dataset: sensor noise causes dramatically more performance degradation than physics perturbations (motor strength, pole length) for this trained policy - heavy sensor noise alone caused 73% degradation, while even large physics deviations caused almost none. This is a genuine, specific result, not an assumed outcome.

Purpose of Dataset

To Support Research Into Predicting Sim-to-real Transfer Risk For Robotic Control Policies Without Requiring Full Physical Deployment Testing, Enabling Faster And Safer Pre-deployment Screening.