A new robotics method cuts the number of required training demonstrations from thousands to one.
Researchers introduced Rational Inverse Reasoning (RIR), a framework that treats imitation learning as an inference problem rather than a pattern-matching one. Instead of memorizing how a human moved, RIR uses a vision-language model to propose compact programs describing why the demonstrator acted that way — goals, subgoals, and constraints encoded in a form a planner can execute. A hierarchical planner then scores those programs by how well they explain the observed behavior, and the system iterates until it converges on the most plausible intent. Tested on a physical Franka FR3 robot arm, RIR succeeded in one-shot and three-shot settings where standard vision-language-model baselines failed.
The performance gap is the headline: RIR beat planning baselines by 34 percentage points in one-shot tasks and 28 points in three-shot tasks when layouts and object sets changed substantially. That matters because layout and object-set shifts are exactly what breaks most deployed robots — the real world does not hold still between training and deployment.
Most imitation learning research attacks the data bottleneck by collecting more data; RIR bets that the bottleneck is abstraction, not volume. It is a reasonable wager, though lab robots on curated tasks have a long history of not surviving contact with a cluttered kitchen counter.