When robots are deployed, they produce huge amounts of data. Engineers need to understand successes and failures to decide what new training data to collect. But the data is hard to interpret, so decisions rely on guesswork.
Behavior graphs compress all robot trial runs into a single directed graph. Nodes group similar behaviors, edges show transitions and outcomes. The graph is built without human labels from the robot's own representations.
In a test with three real-world tasks, 16 of 18 engineers preferred the graph over videos, achieving the same conclusions with much lower perceived time demand and effort. The graph also predicted task progress more accurately than existing methods across many datasets.