ML / AI · Jun 2026
Robust RL Observation Noise Benchmark
A CartPole observation-noise benchmark with a DQN baseline and extensible noise processes, designed for controlled evaluation of reinforcement learning robustness under sensor perturbations. The suite now covers the evaluation loop, DQN agent, and noise processes, with input-validation hardening.
CartPole
Environment
DQN
Baseline
Pluggable
Noise
Problem
Robustness claims for RL policies are often reported under a single noise condition, making cross-paper comparison unreliable. A small, controlled benchmark was needed to characterize degradation across a parameterized family of observation-noise processes.
Approach
I implemented a DQN baseline on CartPole and a noise-injection layer with pluggable processes (Gaussian, biased, drifting), held the training and evaluation protocol fixed, and swept noise intensity to produce degradation curves rather than point estimates.
Results
The benchmark exposes regimes in which the DQN baseline collapses gracefully versus catastrophically and provides a reproducible starting point for evaluating robust-RL methods against a common reference.
Stack