01 / Observe
See what the policy sees.
The scene and wrist views use the same orientation as the model inputs. The task text and measured robot state accompany the images.
SmolVLA / LeRobot / LIBERO
Words into motion.
Every action on record.
A real vision-language-action policy, running in simulation on CPU. Inspect its two camera views, language task and applied actions on one shared timeline.
Language instruction / recorded task
SCENE CAMERAPOLICY INPUT VIEW
WRIST CAMERAPOLICY INPUT VIEW
EPISODE sSIMULATOR s
END-EFFECTOR X / m
END-EFFECTOR Y / m
END-EFFECTOR Z / m
02 / Act
Inspect the executed control.
The trace keeps proposed actions, bounded applied controls, and the observation that produced each action chunk. Terminal observations contain no invented action.
03 / Check
Keep the outcome measurable.
The result comes from LIBERO's task predicate. This is one seeded rollout, with pinned policy and asset revisions, not an aggregate benchmark or a hardware test.