Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Human Imitated Bipedal Locomotion with Frequency Based Gait Generator Network

Yusuf Baran Ates, Omer Morgul

TL;DR

This work tackles robust, human-like bipedal locomotion by marrying a frequency-domain gait generator learned from motion capture with a torque-controlled PPO policy. The gait generator maps morphology and desired speed to smooth joint trajectories via a 32-point FFT, and the PPO controller refines these references under an imitation-plus-gait reward, aided by random state initialization. The method demonstrates improved speed range, slope tolerance, and terrain robustness compared with CMA-ES SIMBICON and standard PPO baselines, while achieving smoother, more human-like joint coordination. This approach offers a practical, data-efficient path to natural locomotion with modest training costs and has potential for extension to full 3D humanoids and richer sensory feedback.

Abstract

Learning human-like, robust bipedal walking remains difficult due to hybrid dynamics and terrain variability. We propose a lightweight framework that combines a gait generator network learned from human motion with Proximal Policy Optimization (PPO) controller for torque control. Despite being trained only on flat or mildly sloped ground, the learned policies generalize to steeper ramps and rough surfaces. Results suggest that pairing spectral motion priors with Deep Reinforcement Learning (DRL) offers a practical path toward natural and robust bipedal locomotion with modest training cost.

Human Imitated Bipedal Locomotion with Frequency Based Gait Generator Network

TL;DR

This work tackles robust, human-like bipedal locomotion by marrying a frequency-domain gait generator learned from motion capture with a torque-controlled PPO policy. The gait generator maps morphology and desired speed to smooth joint trajectories via a 32-point FFT, and the PPO controller refines these references under an imitation-plus-gait reward, aided by random state initialization. The method demonstrates improved speed range, slope tolerance, and terrain robustness compared with CMA-ES SIMBICON and standard PPO baselines, while achieving smoother, more human-like joint coordination. This approach offers a practical, data-efficient path to natural locomotion with modest training costs and has potential for extension to full 3D humanoids and richer sensory feedback.

Abstract

Learning human-like, robust bipedal walking remains difficult due to hybrid dynamics and terrain variability. We propose a lightweight framework that combines a gait generator network learned from human motion with Proximal Policy Optimization (PPO) controller for torque control. Despite being trained only on flat or mildly sloped ground, the learned policies generalize to steeper ramps and rough surfaces. Results suggest that pairing spectral motion priors with Deep Reinforcement Learning (DRL) offers a practical path toward natural and robust bipedal locomotion with modest training cost.

Paper Structure

This paper contains 14 sections, 11 equations, 9 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Bipedal Controller
  3. Data Preparation
  4. Gait Generator Network Training
  5. Simulation Environment and Robot Model
  6. Observation, Action, and Control Scheme
  7. Agent Training
  8. Reward Design
  9. Results
  10. Velocity Comparison Demo
  11. Rotation Demo
  12. Noisy Plane Demo
  13. Velocity Tracking Demo
  14. Conclusion

Figures (9)

  • Figure 1: Controller Loop.
  • Figure 2: Policy Network Architectures
  • Figure 3: Comparison of joint trajectories between our imitation-based controller, ablated variants, and motion-capture reference at $1.2~\text{m/s}$. Each curve represents the mean over three random seeds, and shaded areas indicate $\pm$SD. Motion-capture data correspond to the mean of five trials with speeds between $1.195$ and $1.205~\text{m/s}$.
  • Figure 4: Velocity Comparison Demo Results
  • Figure 5: Average success rate across ramp angles
  • ...and 4 more figures