Search papers, labs, and topics across Lattice.
This paper addresses the challenges of in-hand manipulation by integrating two physics priors鈥攐ne focused on grasp quality and the other on fingertip geometry鈥攊nto a reinforcement learning framework. The global grasp-quality prior enhances the reward structure to promote stable and effective contact distributions, while the local contact-geometry prior optimizes fingertip design for improved rolling performance. Experimental results demonstrate that these priors significantly enhance rotation efficiency, grasp stability, and disturbance rejection in multifingered robotic hands across various tasks, highlighting their effectiveness in improving task robustness and sim-to-real transfer.
Physics-informed priors in reinforcement learning can drastically enhance the robustness of in-hand manipulation, leading to more efficient and stable robotic control.
In-hand manipulation without external sensing is challenging due to uncertainties from finger-object contacts and disturbances by gravity. While reinforcement learning has shown promise in learning complex finger gaiting, existing approaches do not prioritize maintaining well-conditioned grasps for sustained manipulation. We introduce two complementary physics priors for robust in-hand rolling: a global grasp-quality prior derived from classical grasp analysis and a local contact-geometry prior based on fingertip curvature. The grasp-quality prior is used as a dense reward-shaping term that encourages well-distributed contacts with improved worst-case wrench resistance. The contact-geometry prior is expressed in the fingertip geometry that mechanically shapes the contact interface toward task-aligned rolling while reducing off-axis drift. We evaluate the effect of these priors on learning in-hand rolling manipulation for a multifingered robotic hand manipulating three different objects at four palm orientations. Results show significant improvement in rotation efficiency, grasp stability, and disturbance rejection, suggesting that physics priors embedded in both learning and fingertip morphology improve task robustness and sim-to-real transfer. An overview video can be found at https://youtu.be/pdd1wHxQnJM?si=dM-U5kiiPTYsk3Pk.
