DigitCode: Symbolic Tokenization of Hand Motion by Anatomical Units
A new research paper on arXiv (ID: 2608.03127) introduces DigitCode, a method for symbolically representing hand motion by tokenizing it into discrete symbols aligned with anatomical units. The paper, announced as a cross-type submission, addresses the limitation of continuous representations like joint angles or MANO parameters, which, while accurate, lack structure and prevent indexing or editing of individual fingers. Building on Hand Labanotation (HL), which encodes hand motion as a T x 40 grid of direction symbols per bone, DigitCode adapts, groups, and layers HL's alphabet along a hierarchy of units—bone, finger, or whole hand. This approach reduces quantization error by three quarters compared to existing symbolic representations. The key insight is that the choice of anatomical unit, rather than the quantizer family, is the primary lever for improving accuracy. The paper is available at https://arxiv.org/abs/2608.03127.
Key facts
- DigitCode is introduced as a symbolic tokenization method for hand motion.
- It is based on Hand Labanotation (HL), which uses a T x 40 grid of direction symbols per bone.
- DigitCode reduces quantization error by three quarters compared to HL.
- The method adapts, groups, and layers HL's alphabet along a hierarchy of anatomical units: bone, finger, or whole hand.
- The paper argues that the anatomical unit is the key factor in reducing error, not the quantizer family.
- The research is published on arXiv with ID 2608.03127.
- The paper is a cross-type announcement.
- The source URL is https://arxiv.org/abs/2608.03127.
Entities
Institutions
- arXiv