A research team has introduced a novel training approach that enables multi-fingered robot hands to achieve advanced in-hand object manipulation with enhanced safety. The method, detailed in a paper published on arXiv, focuses on teaching a LEAP hand to reach specific object positions and orientations within its grasp through extended sequences of movements. This training recipe, developed entirely within a simulation environment without reliance on human demonstration data, incorporates diverse reset conditions, goal generation from grasp datasets, and an adaptive friction curriculum.The researchers evaluated the policies on a real LEAP robot hand, observing a 95.2% success rate in reaching target object poses within a 2 cm and 0.4 radian margin. A key component of the new approach is a learned residual safety filter, which, when applied, reduced object drops by 75% while maintaining a 91.5% goal-reaching success rate. The safety filter allows for adjustable conservativeness, enabling a balance between task performance and the prevention of drops.The LEAP hand itself is a low-cost, open-source robotic hand developed by researchers at Carnegie Mellon University. It features 16 actuated degrees of freedom and can be assembled for under $2,000, offering an accessible platform for robotics research. Previous research with the LEAP hand has demonstrated its capability for various manipulation tasks, including learning from video data and sim-to-real transfer for tasks like blind in-hand cube rotation.Traditional approaches to dexterous manipulation with multi-fingered hands have often been limited to simpler grasps or basic repetitive motions. As robots undertake more intricate in-hand maneuvers, ensuring safety, particularly preventing objects from being dropped, becomes increasingly important. However, safety metrics are rarely a primary focus in existing research. The new method addresses this by explicitly integrating safety into the learning process.The training process for the LEAP hand policies involved extensive simulation. The policies were designed to transfer directly from simulation to the real robot without further training, using either motion capture or a single wrist-mounted depth camera for perception. This zero-shot transfer capability highlights the fidelity of the simulation and the robustness of the learned policies. The policies proved capable of tasks such as spinning a lightbulb held in an arbitrary pose, reaching out-of-distribution goals, and maintaining stability despite fast arm movements and external disturbances. The use of only a wrist camera also allows for portable skills, such as screwing in a lightbulb. The research team has made the full training recipe available, aiming to provide a platform for further study into performance, reliability, and safety in high-dimensional dexterous robotic systems.