Fade to. Red? Scheduling Coloured Noise Exploration Supports Efficient Robot Learning with Reduced Movement Jerk

This research investigates the use of temporally correlated coloured noise, specifically red noise, for exploration in deep reinforcement learning for continuous control. The study compares various scheduling strategies between white and red noise against generalised state-dependent exploration (gSDE) on four sparse-reward manipulation tasks using a Fetch arm and a Shadow Dexterous Hand. Results indicate that while the final noise colour does not significantly impact learning efficiency or end performance, fading towards red noise reduces movement jerk by half across all tasks compared to white noise. The authors conclude that exploration schedulers fading towards red noise are preferred for off-policy robot learning as they balance learning efficiency, end performance, and smoothness.

Generated from attributed reports · Updated 8 hours ago

There is not enough continuous observation data to show a trend.