On February 7, Xiaomi’s robotics team announced a significant breakthrough in the field of embodied intelligence, introducing their latest achievement, TacRefineNet, on February 5. This innovative framework operates solely based on tactile feedback, eliminating the need for visual input or three-dimensional models of objects to achieve millimeter-level positional adjustments.
This technology aims to address the “last mile” challenge in practical robotic applications, opening up new possibilities for industrial automation. As robotics technology continues to advance rapidly, the importance of tactile perception as a crucial link between intelligent agents and the physical world has become increasingly evident. Just as humans can effortlessly perform delicate tasks like peeling an egg, enhancing robots with similar tactile sensing capabilities is vital for overcoming existing technological limitations.
According to information from IT Home, the standout feature of the TacRefineNet model is its ability to achieve millimeter-level precision in grasping adjustments using only tactile information, without any visual assistance or 3D models of objects. The model is built on high spatial resolution tactile sensors and integrates multi-finger tactile data with proprioceptive information, allowing it to control grasping errors within the millimeter range. In practical tests, whether in simulation environments or real-world scenarios, the model has demonstrated exceptional performance. Notably, it can handle various precision grasping tasks in an automotive factory without needing retraining for specific tasks, truly enabling a plug-and-play experience.
The technical team has validated the performance of TacRefineNet through extensive simulations and real robot testing. Results indicate that the model can quickly adjust diverse initial grasping poses to the target state, reducing average positional errors to the millimeter level. Even with frequently changing object positions and orientations, the system can make precise adjustments through real-time feedback. Remarkably, the model also exhibits generalization capabilities for previously unseen objects, effectively handling unfamiliar items with similar geometric characteristics.
This groundbreaking achievement is the result of a deep integration of data, algorithms, and hardware. On the data front, the research team utilized the MuJoCo physics engine to create a high-fidelity tactile simulator, enabling precise modeling of contact force responses. In terms of algorithms, TacRefineNet employs an end-to-end multimodal learning architecture, combining multi-finger tactile, proprioceptive, and spatial motion information. On the hardware side, the team has integrated high spatial resolution tactile sensors into the fingertips of dexterous hands, with a contact point spacing of just 1.1 millimeters, allowing for the capture of minute surface deformations on objects.
Currently, Xiaomi’s robotics team has released relevant technical details and experimental videos, with more research findings expected to be published in the future.
Original article by NenPower, If reposted, please credit the source: https://nenpower.com/blog/xiaomi-robots-achieve-breakthrough-in-tactile-perception-for-millimeter-level-precision-grasping/
