Rules help frozen robot models handle new wording
In plain words
This paper introduces a method for helping robot models cope with different wordings of the same task.
The authors try many phrasings on a few simulated tasks, use the scores to write a short set of rules, then have a rewriter apply those rules once to each new instruction before the unchanged robot model acts.
On twelve sealed simulated tasks, this improved success over leaving difficult and natural instructions unchanged. The gains were concentrated on tasks outside the model’s training distribution, and the tests do not establish results on physical robots.
The research
If the goal stays the same, why can a robot succeed under one instruction and fail under another, and can evidence from a few tasks help it handle unfamiliar wording?
The paper examines this sensitivity in simulated robot tasks, including a π₀ model that had already been trained with rephrased instructions but remained sensitive.
The method distills scored phrasings of a few tasks into rules that rewrite each new instruction once before the frozen robot model acts.
An oracle phrase search on twelve sealed Bridge tasks showed the in-distribution versus out-of-distribution success gap shrinking from 21 to 3 percentage points, estimating how much wording alone could change results. The authors then applied rules distilled from other tasks to new instructions.
Compared with passing difficult and natural instructions unchanged, Claude’s rule-guided rewriter improved success across those phrase sets. This supports the view that rules can help the frozen model handle new wording in these simulated tasks.
The oracle phrase search estimates wording headroom; it is not the deployable rulebook result. The evaluations covered two robot models in simulation, so they do not establish performance on physical robots or explain why the models are sensitive to wording.
Analysis
Deep
- Authors
- Mikey Watts, Yuchen Cui
- License
- arXiv non-exclusive