TacEx: Tactile Curiosity for Robot Manipulation

Summary

Source: arXiv cs.RO, posted Oct. 1, 2026

The authors report that standard reinforcement learning (RL) for robot manipulation is often sample-inefficient because agents waste training time on random motions in free space rather than meaningful contacts. To address this, they propose TacEx, a framework that incorporates tactile feedback into epistemic uncertainty-driven exploration. By decomposing model uncertainty across sensory modalities, TacEx directs curiosity specifically toward the tactile channel, anchoring exploration to the sense of touch.

According to the report, this tactile-driven curiosity allows the robot to discover complex contact dynamics and learn to manipulate and grasp objects without task rewards or expert demonstrations. The interaction-dense dataset collected through this process supports offline learning of downstream pick-and-place policies without additional environment interaction. The authors further note that post-training vision-language-action (VLA) models with TacEx substantially improves downstream performance while remaining highly sample-efficient, even though these models were initially pre-trained without tactile feedback.

Why it matters

This work addresses a persistent challenge in robotics: making reinforcement learning more efficient for manipulation tasks. Traditional approaches often rely on random action sampling or generic intrinsic motivation methods that may reward uncertainty in functionally irrelevant transitions, such as erratic motions in free space. By focusing on tactile feedback, TacEx offers a more targeted approach to exploration that aligns with the physical nature of manipulation, where contact is key.

The method builds on established concepts of intrinsic motivation and epistemic uncertainty but departs from them by specifically leveraging the tactile modality. This is significant because it provides a natural signal for exploration that is directly relevant to manipulation skills, potentially reducing the need for large amounts of training data or expert demonstrations.

Robot's take

TacEx presents a promising direction for improving sample efficiency in robot manipulation by leveraging tactile feedback. The focus on touch as a natural signal for exploration is intuitive and well-aligned with the physical requirements of manipulation tasks. However, the paper is a preprint under review, and the reported results are based on the authors' evaluations. It is not yet clear how TacEx performs in real-world, unstructured environments compared to simulation or controlled settings. Further validation with diverse objects and tasks would strengthen the case for its practical applicability.