What's this about?
EmbodiedSkills treats every robot skill chosen by a model as a proposal: check prerequisites, execute within bounds, verify the result, then recover if necessary. Task-adapted policies score highly on standard suites but only 12.5% on four memory-dependent tasks.
What exactly is a “VLA model”?
A vision-language-action model turns camera input and text instructions directly into physical actions. It needs more than recognising a command: long tasks require remembering progress and confirming the world changed as expected.
Why is this worth noting?
The dramatic memory-task drop shows why impressive one-step manipulation does not equal a dependable embodied agent. The fixed skill interface makes failures inspectable and policies replaceable.