Tactile Curiosity Drives Robot Interaction
TacEx steers robot exploration toward touch events, not empty-space motion.
The paper argues that tactile feedback is a better target for curiosity-driven reinforcement learning because manipulation skills emerge from contact. Its framework decomposes uncertainty by sensory modality and directs exploration toward the tactile channel. The authors say this produces interaction-heavy data that can support offline pick-and-place learning without task rewards, demonstrations, or more environment interaction during exploration. They also report that post-training vision-language-action models with TacEx improves downstream performance despite those models being pretrained without touch. ArXiv · AI/CL/LG's note
The paper argues that tactile feedback is a better target for curiosity-driven reinforcement learning because manipulation skills emerge from contact. Its framework decomposes uncertainty by sensory modality and directs exploration toward the tactile channel. The authors say this produces interaction-heavy data that can support offline pick-and-place learning without task rewards, demonstrations, or more environment interaction during exploration. They also report that post-training vision-language-action models with TacEx improves downstream performance despite those models being pretrained without touch. ArXiv · AI/CL/LG's note
score 5