EXPERIMENT NOTEBOOK 01 / LIBERO
+From instruction
to robot behavior.
+ Observe what changes when an agent can call a vision-language-action policy.
+ +CURRENT EXPERIMENT
Bowl → plate
Three ways to act
Native simulator outcomesLoading validated results…
+ +Videos show simulation time. Planning and inference pauses are omitted; wall time is reported separately.
+Inside the episode
WHAT THIS RUN SHOWS
+Delegation worked.
Planning benefit remains an open question.
+ B2 inspected the scene and interfaces, then passed the whole task to the VLA in one accepted call. There is no evident task decomposition in this episode. B0 also succeeded. These observations motivate controlled behavior comparisons; they do not establish a general advantage for agent + VLA.
+B1 reached the native control-step limit. Its failure is retained alongside the successful episodes, with the full observable tool trace.
+