Fall Risk-Aware Adaptation Explains Suboptimal Locomotor Performance

Inseung Kang^1,3, Kanishka Mitra², Nidhi Seethapathi^1,2

¹ Department of Brain and Cognitive Sciences, Massachusetts Institute of Technology ² Department of Electrical Engineering and Computer Science, Massachusetts Institute of Technology ³ Department of Mechanical Engineering, Carnegie Mellon University bioRxiv preprint, 2026

bioRxiv Code

Adaptation-aware model of locomotor performance across environmental settings

Safety-aware locomotor learning: inverse adaptation reveals that people do not simply fail to reach an energetic optimum in a novel walking environment; instead, they shift learning rate and symmetry weighting to remain in safer regions of a fall-risk landscape.

Abstract

Human locomotion requires balancing multiple biological objectives, such as metabolic energy efficiency, stability, and symmetry. While models based on optimization successfully predict how humans walk in familiar settings, they fail to explain why individuals adopt inefficient movement patterns in novel environments, even after extensive practice. Here, we show that such suboptimality in a novel environment arises from a fundamental prioritization of safety. We find that individuals do not simply fail to reach an optimal solution; instead, they navigate an environment-dependent risk landscape by mitigating the statistical probability of falling. We find that this risk-averse strategy is explained by adjusting internal learning parameters: specifically, the learning rate and the tradeoff between metabolic cost and symmetry, in a manner that lowers fall risk. To quantify this process, we developed an inverse adaptation modeling framework that works backwards from locomotor performance data to infer the underlying internal learning parameters and how they vary with fall risk. Our analysis reveals that the observed motor performance is explained by a global probabilistic fall risk rather than a local step-based measure of instability.

Adaptation-Aware Model Across Environmental Settings

The paper models split-belt locomotion with a two-level architecture: a low-level feedback controller that maintains stable gait step by step, and a high-level reinforcement learner that updates control policy over time. Two internal parameters are especially important: the learning rate, which controls how aggressively the policy is updated, and the symmetry weight, which determines how much symmetry is prioritized relative to metabolic efficiency. By varying the belt-speed difference, the authors create distinct environmental settings with different theoretical optima and different stability demands.

Figure 2. Adaptation-aware model of locomotor performance across environmental settings, with a low-level feedback controller, a high-level reinforcement learner, and the key learning parameters that govern adaptation.

Inverse Adaptation Framework

To explain the observed behavior, the paper fits time-varying experimental energy trajectories with first-order exponentials and compares them against a library of simulated adaptation profiles spanning different learning rates and symmetry weights. The best-fit parameter pair is the one that minimizes the error between experimental and simulated trajectories. This turns the model into a tool for recovering the internal learning parameters that best explain each participant’s adaptation dynamics.

Figure 3. Inverse adaptation framework for inferring individual-specific learning parameters from experimental locomotor data and simulated adaptation profiles.

Results

The key result is that apparently suboptimal locomotor performance is not well explained by pure energy minimization. As environmental challenge increases, inferred learning rates decrease and symmetry weights increase. The model shows why: higher learning rates reduce energy more aggressively but also raise the probability of falling, especially at larger split-belt speed differences. The inferred parameter regions therefore shift toward safer pockets of the fall-risk landscape, indicating that participants adapt in a fundamentally risk-averse manner.

Figure 4. The learning rate and environmental setting influence locomotor stability, showing representative gait patterns, examples of fall versus no-fall trajectories, the rise of fall probability with learning rate and speed difference, and the higher variance of strides-until-fall as a risk metric.

Safety-Aware Implications for Motor Learning

The broader message of the paper is that slower or energetically suboptimal adaptation in a novel environment need not reflect failure. Instead, it can reflect a strategic prioritization of safety. By framing locomotor adaptation in terms of a fall-risk landscape, this work offers a useful lens for understanding how people learn in destabilizing environments and for designing rehabilitation technologies that account for safety, symmetry, and energy together rather than treating them as isolated objectives.

Fall risk landscape across environmental settings

Figure 5. Fall risk landscapes across environmental settings, showing that the inferred learning parameters shift toward lower-risk regions as the environment becomes more challenging.

BibTeX

@article{kang2026fall,
  title={Fall risk-aware adaptation explains suboptimal locomotor performance},
  author={Kang, Inseung and Mitra, Kanishka and Seethapathi, Nidhi},
  journal={bioRxiv},
  pages={2026--03},
  year={2026},
  publisher={Cold Spring Harbor Laboratory}
}