In a flow-matching VLA, single-frame proprioceptive state works best in the VLM prefix, while 8-frame state history works best in the action prefix, with gains from genuine temporal content rather than extra tokens.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.RO 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
How Should Vision-Language-Action Models Use Proprioceptive State?
In a flow-matching VLA, single-frame proprioceptive state works best in the VLM prefix, while 8-frame state history works best in the action prefix, with gains from genuine temporal content rather than extra tokens.