Event
NavBind: Diagnosing Navigation VLA Instruction-Action Binding
1 reports · 1 sources
Josh Qixuan Sun and other researchers proposed the NavBind method to diagnose whether navigation VLA models truly follow instructions: under a fixed navigation context, it checks whether changing the instruction affects the probability distribution of actions. The method constructs 3,252 route-fork contexts by recombining route-aligned instruction segments from R2R-CE and RxR-CE; each context pairs two complete instructions that share the description of the route already traveled but specify different branches, and evaluation is performed under the same agent state and navigation history. The authors state that, among the evaluated models, higher navigation success rates do not imply that the models follow instructions more faithfully.
Report timeline
- NavBind Diagnoses Instruction-Action Binding at Route Forks in Navigation VLA Models
NavBind proposes a method for diagnosing whether navigation vision-language-action (VLA) models truly follow instructions: with the navigation context fixed, it checks whether changing the instruction affects the probability distribution of actions. The authors recombined route-aligned instruction segments from R2R-CE and RxR-CE to construct 3,252 route-fork contexts, each pairing two full instructions that share a description of the route already traveled but specify different branches, and evaluated them under the same agent state and navigation history.
arXiv cs.RO ·