Robot Safety Commands: MIT Robot Widens One Warning to Electronics; GPT-4o Errs Up to 9 of 105 Tests
Reported by Tech Xplore and Tech Times2 sources ↓
MIT and NYU researchers developed FLEX, which uses GPT-4o to turn a typed correction into broader robot-avoidance rules. Tech Times reports each of its three editing operations made errors in 3.8% to 8.6% of isolated trials.
Key points
- FLEX scored 0.99 normalized reward in a navigation test requiring all three editing operations
- Comparison methods scored 0.83, 0.78 and 0.69 in that test
- Add, Generalize and Mask each made errors in 9, 6 and 4 of 105 trials, respectively
- The paper describes Mask as GPT-4o scoring feature relevance; MIT News describes a computer-vision step
- The authors say the method assumes quasi-static rigid objects and does not account for changes in preference strength
Sources
Details
Tech Times reports FLEX represents a robot's reward as a weighted sum of features. GPT-4o edits the feature list after a human correction using Add, Generalize and Mask; demonstrations determine the weights. The paper tested the system in seven 9-by-9 grid navigation scenarios and simulated five-stage tasks using a seven-degree-of-freedom Franka arm.
In a navigation test requiring all three operations, FLEX scored 0.99 normalized reward, with a standard error of 0.01 across five random seeds. The comparison methods scored 0.83 for mask-only editing, 0.78 for a single unscaffolded language-model query and 0.69 for add-only editing. The appendix reports that Add made errors in 9 of 105 trials, Generalize in 6 and Mask in 4; errors included incorrect edits and edits that were omitted.
The accounts differ on how Mask works. MIT News describes computer vision removing features for objects no longer present, while the paper, as reported by Tech Times, describes GPT-4o scoring feature relevance and does not describe a visual step. MIT News says the process takes about five seconds, but Tech Times says the paper does not confirm that runtime.
The authors list assumptions and limits including quasi-static rigid objects, possible failures of language-model commonsense for personal or counterintuitive preferences, and ignoring corrections that change the strength of a preference. Tech Times also reports that avoidance is a soft reward term, not a guaranteed safety constraint, and that the paper makes no safety guarantees.
Written from coverage as of
Images
- Tech Xplore
- Tech Times