The abstract for a new arXiv paper, "When Do Discrete Actions Make Good Continuous Controllers?" (arXiv:2610.20429), takes aim at a common assumption in robotics. It notes that while modern AI models like LLMs predict discrete tokens, today's robotic foundation models represent actions as continuous vectors. The authors then say they "challenge this common consensus" by asking whether continuous action parametrizations are truly needed.
The framing draws an implicit parallel to the discrete token paradigm that has proven powerful in language modeling. The question is whether a similar discrete representation could work for robot control, potentially simplifying action spaces or improving interpretability. However, the abstract is truncated and does not include the paper's findings or conclusions.
As a result, the source only establishes the motivation and research question. It does not provide evidence for or against discrete actions, nor does it compare specific methods. Readers interested in the outcome will need to consult the full paper.