New method captures individual motor quirks using AI to power realistic digital avatars

Scientists have developed a machine-learning approach to automatically learn and replicate the distinctive movement patterns that make each person's motion signature unique. The technique could improve virtual avatars and robots designed to work alongside humans in settings ranging from physical therapy to sports training.

Why individual movement patterns matter#

Each person moves in a subtly different way—a distinctive velocity signature that remains consistent over time but varies from individual to individual. While this variability is rich and nuanced, most computational models of human motion flatten it into simplified, generic descriptions. This gap between reality and current models creates friction when designing autonomous agents and robots that need to mimic or work alongside humans in group activities like rehabilitation, athletics, and manufacturing environments.

The approach: Learning motion from data#

The team proposes a fully data-driven method built around long short-term memory (LSTM) neural networks—a type of AI architecture well-suited to capturing patterns in sequential data over time. Rather than hand-crafting rules about how humans move, their system learns directly from recorded motion data, focusing on how motion amplitude (the magnitude of movement) provides a useful way to characterize each person's distinctive motor behavior.

Validation and results#

Researchers tested the model using real human recordings of spontaneous oscillatory motion—repetitive, natural movements like swaying or rhythmic limb motion. Statistical analysis confirmed that the trained network successfully reproduces both the velocity distribution and amplitude patterns of the individual it learned from, while remaining distinct from those of other people. The method thus captures what makes one person's movement signature genuinely unique, rather than producing generic or blended motion.