Poster #161 - Walking and falling: Using simulated robots to model variability and error in the development of infant walking

• In Event: 3-003 - Poster Session 09
  In Poster Session: PS 09 Section - Perceptual, Sensory, Motor

Integrative Statement

Walking is a highly creative process. Going from one place to another involves more than merely alternating the legs. To be functional, walking movements must be continually modified to suit variations in local conditions. Moreover, walking without falling is a formidable challenge—involving perceptual guidance and balance control. Despite recent advances in artificial intelligence, roboticists have not yet created robots that learn to walk as proficiently and adaptively as do human infants.

What is the optimal training for functional walking? Previous work indicates that infants’ natural training regimen is immense and highly varied (Adolph et al., 2012; Lee et al., 2018). New walkers take thousands of steps per day. They step in every direction, walk along curved and winding paths, and long jaunts are punctuated by frequent starts and stops (Figure 1A). New walkers also fall dozens of times per day. But the penalty for error is low (unless infants fall from a height). After falling, infants rarely cry, and they are back at play within a few seconds (Han et al., 2018). Despite a new wealth of descriptive data about infants’ everyday walking experiences, it is not clear whether infants’ natural training regimen is in fact a good way to learn to walk. Instead, infants might learn betterif their initial training input were less varied and carried a higher, more salient penalty for error, that is, if their training more closely resembled that of robots and therapy for disabled children.

Because controlled rearing studies are not feasible with human infants, we used simulated robots as an embodied model system to test whether varied walking paths and a low penalty for error are beneficial or detrimental for learning to walk. We trained simulated robots to walk using three geometric walking paths (line, circle, square; Figure 1B), and five varied infant-generated walking paths (compiled from 90 infants during free-play). The varied walking paths were ranked based on the relative variability of their features (i-v; Figure 1C). In addition, we manipulated the penalty for falling during training (using low and high penalties). 1000 robots were trained per path and penalty value.

Findings revealed that varied input with a low penalty for error is an optimal training regimen. The varied infant-based training regimens promoted more generalization than the geometric training regimens (Figure 2A). Moreover, performance on new paths improved as the variability within infant-based training paths increased (Figure 2B). Training with a low penalty for falling also improved walking performance (Figure 2C). Importantly, this benefit was greater for un-trained paths than for trained paths (Figure 2D), suggesting that a low penalty for error during training promotes generalization. We propose that high variability and a low penalty for error are natural features of infant walking that play crucial, functional roles in developing real-world walking skill. Thus, infants’ everyday walking experience constitutes a natural training curriculum that facilitates learning and generalization. Our findings provide insights for training adaptive motor skills (like walking), and may inform theoretical, empirical, and clinical work on motor learning.

Authors

• Justine Hoch, New York University
  Presenting Author

• Ori Ossmy, New York University
  Non-Presenting Author

• Patrick MacAlpine, University of Texas at Austin
  Non-Presenting Author

• Shohan Hasan, New York University
  Non-Presenting Author

• Peter Stone, University of Texas at Austin
  Non-Presenting Author

• Karen E Adolph, New York University
  Non-Presenting Author