Robot Learning

Gradient Clipping

Gradient clipping is a training technique that limits the magnitude of gradients before a parameter update, most commonly by rescaling the global gradient norm to a threshold, preventing rare large gradients from destabilizing optimization. Originally motivated by exploding gradients in recurrent networks, it is now near-universal in training large transformers and reinforcement learning systems, where nonstationary data and value bootstrapping produce heavy-tailed gradient noise.

Why it matters for physical AI

Stable optimization is a precondition for the long, expensive training runs behind robot foundation models, and clipping is one of the small set of defaults that keep large-scale policy training from diverging.