Action Tokenization Bridges Language Models and Robot Control in VLA Systems

Robots act on the physical world through motor commands, which do not naturally match the token-based representations used by foundation models. Action representation serves as the interface linking these two computational domains, since an action's meaning depends on the robot, the task, its physical state, and execution timing. Vision-Language-Action models must convert multimodal inputs into executable motor behavior, which requires a representation of actions that neural architectures can process efficiently. Continuous control vectors preserve precision but integrate poorly with token-based models, while discretized action tokens allow autoregressive prediction at the cost of quantization error. The chosen representation also shapes how easily a model learns, trading fine motor detail against the length of planning horizons.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.



Discussion (0)
Log in to join the discussion and vote.
Log in