DexAnyTwist: Learning General Dexterous Twisting with Hybrid Manipulation System Identification

Developing general-purpose dexterous manipulation, particularly for intricate twisting tasks, remains a long-standing challenge. Large-scale training on massive datasets encompassing a vast spectrum of geometric and physical properties is a promising path. Yet, the hybrid dynamical space inherent in heterogeneous datasets creates intrinsic conflicts, where standard learning paradigms fail to synthesize a cohesive control strategy across disparate physical laws. We present DexAnyTwist, a framework that reformulates the general twisting task as a problem of identification and control within a hybrid dynamical system. This approach utilizes forward partitioning to isolate data modes based on expert performance and backward refinement to optimize specialized policies. A learned gating mechanism then dynamically composes these experts, effectively mitigating interference between distinct dynamic modes. Beyond achieving state-of-the-art performance on our large-scale dataset, our framework spontaneously evolves manipulation primitives consistent with human twisting, and demonstrates superior zero-shot generalization to novel real-world geometries. This study not only solves a complex class of twisting tasks but also establishes a scalable pathway for robots to acquire general-purpose contact-rich manipulation skills.