Infant Locomotor Development: Spatial and Temporal Variability Dissociation

Comparing Human and AI Locomotion

• 💡 The central research question explores whether current AI models learn to move in the same way as humans.
• 🧠 The study employs a predictive approach, characterizing canonical human behavioral patterns and comparing them to AI models on learning time scales.

Infant Locomotor Data Collection

• 🔬 Vicon motion capture system was used to collect movement data from 19 children (ages 9-45 months) and 9 adults.
• ✅ Key gait metrics extracted included stride length (spatial), stride frequency (temporal), and speed, all normalized by the child's height.

Spatial and Temporal Dissociation in Children

• 🎯 Children exhibit a surprising dissociation between spatial and temporal variability in their locomotor patterns.
• 📈 They show an adult-level of spatial coordination, with a strong speed-dependent stride length relationship.
• 📉 However, children demonstrate a weaker speed-dependent stride frequency relationship compared to adults, suggesting less developed temporal coordination.
• 🔑 This observed dissociation holds consistently at the individual level, not just as a group average.

Limitations of Current AI Models

• 🤖 Vanilla Reinforcement Learning (RL) agents, using typical reward functions (e.g., tracking forward velocity, maintaining torso height, conserving energy), do not replicate human-like locomotor development.
• ❌ As RL agents train, the relationship between their stride frequency and speed actually weakens, which is the opposite of what is observed in human adults.

Hypotheses for Future AI Development

• 💡 Future models could incorporate learning to stand before walking to improve postural control and potentially influence locomotor patterns.
• 🧠 Inspiration from neuroscience, such as central pattern generators for rhythmic aspects and reflexes for spatial corrections, might be integrated.
• 🌱 Other hypotheses include adapting to a changing body on learning time scales or using different cost functions for temporal versus spatial parameters.