Tactile Curiosity Drives Robot Interaction
Abstract Overview
This paper introduces TacEx, a reinforcement learning exploration framework that decomposes epistemic uncertainty by sensory modality and explicitly weights tactile uncertainty to guide robot behavior toward contact-rich interactions. The method combines visual, tactile force-map, and latent state representations in an ensemble dynamics model, then uses modality-specific disagreement as an intrinsic exploration signal. In simulation, the authors study both reward-free exploration from scratch and post-training of frozen vision-language-action policies through diffusion steering on contact-rich manipulation tasks. The reported evidence indicates that emphasizing tactile uncertainty increases meaningful object interaction during exploration and yields data that can support downstream offline pick-and-place learning without further environment interaction. The experiments also show a tradeoff between focused tactile exploration and broader multi-object coverage when visual uncertainty is added.
Novelty
The main novelty is a modality-weighted uncertainty objective for exploration that explicitly prioritizes tactile predictive uncertainty instead of treating epistemic uncertainty as a single undifferentiated scalar over the full state. The work is also distinctive in applying this tactile curiosity principle both to reward-free manipulation data collection and to sample-efficient post-training of frozen vision-language-action models.
Results
In reward-free manipulation experiments, variants that include tactile disagreement produce more frequent object interaction and grasp attempts than variants without tactile disagreement, and the resulting datasets support stronger offline pick-and-place performance. In the multi-object setting, tactile plus visual disagreement increases interaction diversity, while tactile-only exploration is more focused. For diffusion steering on contact-rich LIBERO-90 tasks, performance gains are task-dependent, but the best-performing curves typically include tactile disagreement, and the appendix reports the tactile+vision variant as achieving the highest mean final return among compared bonuses on a five-task analysis.
Key Points
- TacEx steers exploration by decomposing uncertainty across visual, tactile, and latent channels and assigning explicit weight to tactile disagreement.
- Tactile-weighted reward-free exploration leads to more contact-rich behavior and produces offline data that improves downstream sparse-reward pick-and-place learning.
- When used for post-training frozen VLA policies via diffusion steering, tactile uncertainty is a useful exploration signal, with visual uncertainty providing complementary gains in multi-object diversity and some downstream tasks.