A new robot-learning framework skips random flailing and lets curiosity about touch, not generic uncertainty, guide exploration instead.
Researchers built TacEx, a reinforcement learning framework that directs a robot's intrinsic curiosity toward tactile feedback rather than treating all sensory uncertainty the same way. Standard RL exploration relies on random action sampling or broad uncertainty-based curiosity, both of which waste training time on motions in free space that never make contact with anything. TacEx instead decomposes a robot's model uncertainty by sensory channel and specifically rewards it for seeking out tactile surprises, pushing it toward contact-rich behavior like grasping and pushing. Without any task-specific reward or human demonstrations, the robot builds a dataset dense with physical interactions, which the researchers then used to train pick-and-place policies offline.
Robot manipulation has long been bottlenecked by sample inefficiency, since RL agents typically burn most of their training budget moving through empty space rather than learning what happens on contact. Narrowing curiosity to touch specifically, instead of rewarding uncertainty anywhere in the model, is a more targeted fix than prior epistemic-uncertainty methods, which still chase noise in irrelevant free-space motion.
The more notable result may be the post-training payoff: vision-language-action models that were never trained on tactile data in the first place got a real performance boost after post-training on TacEx's touch-driven data, and did so with a small amount of additional interaction rather than a full retraining run.