Title: ContactExplorer: Contact Coverage-Guided Exploration for General-Purpose Dexterous Manipulation

Abstract:

While reinforcement learning has demonstrated significant prowess in areas like Atari gaming, navigation, and locomotion—where exploration is frequently driven by novelty in states or dynamics—dexterous manipulation presents a distinct challenge. These tasks demand complex physical interactions between hands and objects, yet current approaches often struggle with unstable novelty signals based on contact, inefficient distance-based novelty metrics, or an over-reliance on task-specific priors. To address these limitations, we introduce ContactExplorer, a novel, general-purpose exploration framework designed specifically for dexterous manipulation.

ContactExplorer conceptualizes contact as the intersection between keypoints on the hand and points on the object’s surface. This approach encourages dexterous hands to uncover a wide variety of novel contact configurations, specifically focusing on which fingers interact with which specific regions of the object. The method employs a contact counter that is conditioned on discretized object states, which are derived using learned hash codes. This counter tracks the frequency of interactions between individual fingers and various parts of the object. We utilize this data in two synergistic manners: first, by assigning a count-based contact coverage reward that incentivizes the discovery of new contact patterns; and second, by implementing an energy-based reaching reward that steers the agent toward contact areas that have been less explored.

We tested ContactExplorer across a broad spectrum of dexterous manipulation tasks. Our experiments indicate that ContactExplorer significantly enhances both sample efficiency and success rates compared to existing exploration strategies. Furthermore, the contact patterns acquired through ContactExplorer demonstrate robust transferability to real-world environments. For more details, visit the project page at https://contact-explorer.github.io.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC