Biomechanical Comparisons Reveal Divergence of Human and Humanoid Gaits

Luying Feng, Yaochu Jin, Hanze Hu, Wei Yang ChenarXiv 2026 · arXiv

Abstract

It remains challenging to achieve human-like locomotion in legged robots due to fundamental discrepancies between biological and mechanical structures. Although imitation learning has emerged as a promising approach for generating natural robotic movements, simply replicating joint angle trajectories fails to capture the underlying principles of human motion. This study proposes a Gait Divergence Analysis Framework (GDAF), a unified biomechanical evaluation framework that systematically quantifies kinematic and kinetic discrepancies between humans and bipedal robots. We apply GDAF to systematically compare human and humanoid locomotion across 28 walking speeds. To enable reproducible analysis, we collect and release a speed-continuous humanoid locomotion dataset from a state-of-the-art humanoid controller. We further provide an open-source implementation of GDAF, including analysis, visualization, and MuJoCo-based tools, enabling quantitative, interpretable, and reproducible biomechanical analysis of humanoid locomotion. Results demonstrate that despite visually human-like motion generated by modern humanoid controllers, significant biomechanical divergence persists across speeds. Robots exhibit systematic deviations in gait symmetry, energy distribution, and joint coordination, indicating that substantial room remains for improving the biomechanical fidelity and energetic efficiency of humanoid locomotion.

Introduction

As humanoid robots move from laboratories into real-world environments, "walking well" has become a priority beyond mere stability. Effective locomotion is essential for energy efficiency, social acceptance, and hardware longevity. Currently, many humanoid controllers are trained using Reinforcement Learning (RL) or Imitation Learning (IL) to mimic human joint kinematics. However, looking like a human does not necessarily mean the robot is moving like one; there remains a significant gap in understanding the underlying forces and energy patterns (kinetics). Furthermore, most studies evaluate walking at a single "preferred" speed, ignoring how gait dynamics evolve across a continuous range of speeds.

The Gait Divergence Analysis Framework (GDAF)

To address these gaps, we introduced GDAF, a standardized framework that evaluates gait through three primary dimensions:

  1. Waveform Similarity: Utilizing Pearson correlation to assess how closely the shapes of joint angle, moment (torque), and power curves match human data.
  2. Bilateral Symmetry: Measuring the "Symmetry Index" (SI) to determine if the robot’s left and right legs perform identically—a hallmark of healthy human gait.
  3. Energetic Behavior: Analyzing the positive and negative work performed by each joint.

The framework culminates in a Comprehensive GDAF Cost ($C_v^{GDAF}$), a single metric that aggregates these factors to provide a quantitative score for gait quality at any given speed.

Experimental Methodology

The study utilized two primary data sources to benchmark performance:

  • Human Data: A large-scale open-source biomechanical database of 22 able-bodied adults walking on treadmills.
  • Humanoid Data: A newly collected, speed-continuous dataset from a Unitree G1 humanoid robot. This dataset covers 28 discrete speeds ranging from 0.5 m/s to 1.85 m/s in fine-grained 0.05 m/s increments.

Biomechanical Divergence Analysis

A primary finding of our analysis is the discovery of a "knee-dominant" power strategy in the Unitree G1. While humans generate significant power through ankle "push-off," the robot generates substantially more positive work at the knee and significantly less at the ankle, as shown in the waveform comparisons below.

Comparison of joint angles, moments, and power between human and robot across different speeds.

Qualitative analysis of torque-angle loops further highlights these discrepancies. Human patterns are smooth and self-similar across speeds, whereas the robot's loops appear distorted and vary irregularly. This indicates a lack of consistent "energy storage-release" patterns, which are critical for efficient biological locomotion.

Torque-angle loops for hip, knee, and ankle joints comparing human and robot dynamics.

Key Insights

  • The Slow-Speed Bottleneck: Counter-intuitively, the robot’s gait quality improves as it moves faster. At slow speeds, the controller struggles to maintain symmetry and smooth energy modulation. The Comprehensive GDAF Cost improved significantly as speed increased, dropping from 0.5106 at 0.5 m/s to 0.2941 at 1.5 m/s.
  • Proximal-to-Distal Gradient: The robot matches human motion effectively at the hips and knees (proximal joints), but similarity breaks down at the ankles and feet (distal joints). This is likely due to the robot's rigid actuators compared to the complex, elastic tendon structures in human feet.

Discussion and Future Directions

The GDAF framework reveals that even when robots appear human-like, they rely on fundamentally different mechanical strategies. The robot’s rigid design currently limits its ability to replicate the passive energy dynamics of human tendons and muscles. Future work will extend GDAF to more complex tasks like stair climbing and integrate the GDAF cost function directly into RL reward functions. This would allow robots to "self-correct" their gaits toward more human-like, energetically efficient patterns during training, bridging the gap between biological and mechanical locomotion.

Summary updated

Abstract

It remains challenging to achieve human-like locomotion in legged robots due to fundamental discrepancies between biological and mechanical structures. Although imitation learning has emerged as a promising approach for generating natural robotic movements, simply replicating joint angle trajectories fails to capture the underlying principles of human motion. This study proposes a Gait Divergence Analysis Framework (GDAF), a unified biomechanical evaluation framework that systematically quantifies kinematic and kinetic discrepancies between humans and bipedal robots. We apply GDAF to systematically compare human and humanoid locomotion across 28 walking speeds. To enable reproducible analysis, we collect and release a speed-continuous humanoid locomotion dataset from a state-of-the-art humanoid controller. We further provide an open-source implementation of GDAF, including analysis, visualization, and MuJoCo-based tools, enabling quantitative, interpretable, and reproducible biomechanical analysis of humanoid locomotion. Results demonstrate that despite visually human-like motion generated by modern humanoid controllers, significant biomechanical divergence persists across speeds. Robots exhibit systematic deviations in gait symmetry, energy distribution, and joint coordination, indicating that substantial room remains for improving the biomechanical fidelity and energetic efficiency of humanoid locomotion. This work provides a quantitative benchmark for evaluating humanoid locomotion and offers data and versatile tools to support the development of more human-like and energetically efficient locomotion controllers. The data and code will be made publicly available upon acceptance of the paper.
BibTeX
@misc{feng2026biomechanical,
  author = {Luying Feng and Yaochu Jin and Hanze Hu and Wei Yang Chen},
  title = {{Biomechanical Comparisons Reveal Divergence of Human and Humanoid Gaits}},
  year = {2026},
  eprint = {2602.21666},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2602.21666},
  howpublished = {arXiv},
}
Figure