UniDexTok: A Unified Dexterous Hand Tokenizer from Real Data

UniDexTok introduces a unified tokenization system that standardizes how different dexterous robotic hands represent their states, enabling cross-embodiment learning from real-world data. By mapping diverse hand kinematics to a shared 22-degree-of-freedom interface, the system achieves sub-millimeter reconstruction accuracy—a 99% improvement over previous approaches—while eliminating the need for simulation or manual retargeting.

UniDexTok addresses a fundamental fragmentation problem in robotic manipulation research: the lack of standardized representations across different dexterous hand designs. While parallel grippers have converged toward similar mechanical principles, dexterous hands vary dramatically in joint configuration, kinematics, and degrees of freedom. This heterogeneity has historically isolated datasets to specific embodiments, limiting training data and preventing the emergence of generalizable manipulation models.

The breakthrough lies in UDHM's semantic mapping layer, which translates raw joint states from any dexterous hand—human or robotic—into a unified 22-DoF representation. UniDexTok then learns discrete tokens conditioned on embodiment identity, enabling cross-embodiment knowledge transfer. The experimental results demonstrate transformative performance: reducing mean per-joint angular error from 15.63 degrees to 0.16 degrees represents a shift from centimeter-scale to sub-millimeter precision.

For the robotics and AI communities, this work accelerates the path toward generalist manipulation models. By pooling previously siloed datasets into a common representational space, researchers can train on aggregate data volumes that individual embodiments cannot provide alone. The zero-shot and few-shot capabilities suggest emerging embodiments require minimal additional training.

This standardization framework positions tokenized dexterous hands as foundational infrastructure for future multimodal AI systems combining vision, language, and tactile understanding. The approach mirrors successful paradigms in computer vision, where standardized representations enabled rapid scaling of model capabilities. Expect downstream adoption across industrial robotics, humanoid platforms, and embodied AI research programs seeking interoperable manipulation solutions.

• →UniDexTok reduces hand reconstruction error by 99%, achieving sub-millimeter accuracy through unified tokenization across heterogeneous dexterous hand designs.
• →A shared 22-DoF semantic interface enables cross-embodiment learning, allowing data from multiple hand types to improve individual embodiment performance.
• →Eliminating retargeting and simulation dependencies reduces computational overhead while preserving real-world data fidelity for training robust manipulation models.
• →Zero-shot and few-shot generalization capabilities suggest the framework scales efficiently to novel dexterous hand designs with minimal additional training.
• →Unified hand representations establish standardized infrastructure for integrating dexterous manipulation into multimodal AI systems and embodied agents.