Event

Study: Discrete Action Policies Can Rival Continuous Policies

1 reports · 1 sources

Zeji Yi and co-authors published a comparative study challenging the common view that "continuous action parameterization is superior to discrete." They compared continuous and discrete policies' single-step regression and multi-step flow variants in single-task regression, multi-task VLA training, and both simulation and real-world settings, finding that even with the most naive uniform binning quantization, discrete action prediction can match or even exceed continuous action policies in task success rate. The authors state that the difficulty of discrete action policies lies not in action precision, but in the decomposition error caused by predicting the product distribution across time and action dimensions; explicitly injecting geometric information into discrete policies by, for example, mimicking the noising process of continuous flow matching also did not bring consistent improvement.