Robot planners and action policies: what does each output?

SayCan selects a skill

SayCan combines probabilities from a language model and a value function to select a skill. A language model scores whether a skill would help satisfy an instruction. A value function estimates whether that skill can succeed from the current state. Their combination selects a skill for the robot to execute. The selection process repeats, retaining the selected steps in its context, until it chooses to finish.

Here, the planning decision is which available skill should come next. A plausible sequence of task names alone would leave out the feasibility estimate that SayCan uses to ground its choice.

OpenVLA predicts an action representation

The original OpenVLA paper describes a different interface. An observation image and a language instruction enter the model; its output represents a robot control action. The authors map continuous action values to discrete tokens so that a language-model backbone can predict them.

The tokens therefore carry action information. They are not simply a written explanation of what a robot might do. The paper describes seven-dimensional control actions, rather than the skill selection used in the SayCan comparison.

Read the handoff before comparing the models

SayCan selects a skill for the robot to execute, while OpenVLA predicts tokens representing a robot control action. These are separate research systems; this comparison does not connect them into one robot or rank their performance.