A regression robot policy reportedly rivals generative baselines with faster inference

In an October 2026 paper, researchers say standard mean-squared-error training produces larger gradients for observations with large action-prediction errors. Their alternative, HT-Policies, learns how error scales vary by input and predicts actions in one feed-forward pass. They report success rates competitive with generative policy baselines across four simulation benchmarks and real-robot evaluations, with faster training and inference.