Researchers have a new way to stop a robot from handing you a knife blade-first — without rebuilding the entire model.
Robot policies trained on large, varied demonstration datasets often pick up multiple behavior modes, some of them unsafe. The standard fixes are expensive: curating the original data and retraining from scratch, or bolting on inference-time steering that slows every decision the robot makes. A paper published today introduces MoRE (Mode Redirection), a short "uncloning" step that trains a temporary classifier to identify undesired modes, then distills that signal directly into the policy weights. A retain loss keeps the robot competent at its desired tasks. Once done, the corrected policy runs standalone — no extra compute at deployment.
The gap this closes is real. Inference-time steering adds overhead that compounds across thousands of robot decisions per hour; full retraining requires keeping the original demonstration dataset around and burning the compute to use it. MoRE tested across eight simulated and real-world tasks, lifting average deployment success rates by 44 percentage points over uncorrected mixed-mode policies, and matched the performance of filtered-data retraining without the data access requirement.
The technique generalizes across Diffusion Policy and Pi0.5 VLA backbones, which matters: those are two of the more widely adopted robot policy architectures right now. Whether this holds as demonstration datasets get larger and behavior modes get subtler is the obvious next question — but a 44-point jump in success rate is a hard number to dismiss.