BEE, a new training method for robot AI, learns how much to trust a human's mid-task correction instead of copying it exactly.
Researchers built BEE for vision-language-action models, the AI systems that let robots handle long, multi-step manipulation tasks like assembling parts or sorting objects. The problem is that a single sloppy human correction during training can undo a robot's progress, especially in the split-second moments where precision decides success or failure. BEE's fix is a "Correction Model" that predicts how consistent a human's correction is likely to be along each dimension of an action, staying close to the human where people reliably agree and loosening up where they don't. The team tested BEE on three real-world manipulation tasks and one simulated LIBERO-Pro task, matching the training data budget used by two rival methods, RLT and DSRL.
BEE won on every task, averaging 91.2% success across those four benchmarks - three real robots plus one simulation - against 57.5% for RLT and 42.1% for DSRL. It also needed the least human intervention on every real-world task, which matters because that hands-on correcting is the actual bottleneck in physical robot training: a person standing there nudging the arm back on track.
One of those four wins came from a simulation, not a real robot, so treat the headline number as a mixed bag. Still, cutting how often a human has to intervene is the metric that decides whether work like this ever reaches a warehouse floor.