A framework to quantify and fix 'task steerability,' the common failure of robots to respond to new instructions while mid-task.
arXiv · March 19, 2026 · 2603.17300
The Takeaway
It identifies that multi-task pretraining often creates rigid behaviors and provides a self-refinement pipeline to improve interactivity. This is essential for deploying robots in dynamic environments where users need to change instructions on the fly.
From the abstract
Despite strong multi-task pretraining, existing policies often exhibit poor task steerability. For example, a robot may fail to respond to a new instruction ``put the bowl in the sink" when moving towards the oven, executing ``close the oven", even though it can complete both tasks when executed separately. We propose ReSteer, a framework to quantify and improve task steerability in multitask robot policies. We conduct an exhaustive evaluation of state-of-the-art policies, revealing a common lac