Manipulation
-
PAPERNEO-VLA drops the vision encoder so action policies can read visual features at any depth
A paper from MMLab@NTU argues that vision-language-action models lose action-critical detail in their visual encoders and proposes NEO-VLA, an encoder-free design.