OpenDIM Lab studies how robots can sense, learn, and act in the physical world. Our research combines tactile sensing, multimodal perception, robot learning, and dexterous manipulation to develop machines that can interact with objects reliably and precisely.
We focus on three closely connected research directions:
We develop vision-based tactile sensors and learning-based perception methods for understanding contact, geometry, material properties, and motion. Tactile observations provide information that is often unavailable to cameras alone, especially when objects are occluded or when successful manipulation depends on subtle contact changes.
Our work explores tactile representations, contact localization, surface reconstruction, tactile motion estimation, and multimodal fusion. By combining touch with vision, proprioception, and other sensory signals, we aim to give robots a more complete understanding of their physical interactions and enable robust perception under changing lighting, viewpoints, and contact conditions.
We build scalable tactile simulation and sim-to-real learning methods that transfer robot skills from virtual environments to the real world. Simulation enables large-scale data generation, controlled experimentation, and systematic evaluation of sensing and control strategies that would be expensive or slow to collect entirely on physical robots.
Our research investigates differentiable and data-driven tactile simulation, reinforcement learning, imitation learning, and multimodal policy learning. We design training pipelines that reduce the gap between simulated and real tactile observations, allowing robots to learn useful behaviors with less physical data and improved generalization to new objects, surfaces, and interaction conditions.
We develop touch-guided skills for in-hand manipulation, bimanual coordination, and precise contact-rich tasks. These tasks require robots to continuously reason about contact forces, object motion, hand configuration, and uncertainty while coordinating many degrees of freedom.
Our goal is to create manipulation policies that are both capable and reliable, including skills for grasp adjustment, object rotation, tactile servoing, pushing, reorientation, and tool-like interaction. By integrating tactile feedback with learned models and closed-loop control, we seek to enable dexterous robots to adapt their actions online rather than relying only on precomputed trajectories or precise object models.