.png){kind=icon}
.png){kind=icon}
Return
A Unified Reinforcement Learning Framework for Payload-Aware Humanoid Control With Integrated Inertial and State Estimation
.png){kind=icon}
PRE
AI
S
Z
C
J
H
DOI：10.1109/tmech.2025.3648874.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
This article presents an integrated reinforcement learning framework for payload-aware humanoid control that enables online adaptation to varying loads. Unlike conventional modular designs, our approach employs a single proprioception-driven policy network that unifies three core functions–motion control, inertial parameter identification, and state estimation–within a cohesive architecture, thereby avoiding inter-module coupling and simplifying the overall system. To support accurate inertial identification across diverse gait patterns, we design parameterized gait templates based on phase-locked oscillators and polynomial interpolation, providing structured excitation while ensuring kinematic feasibility. The control policy is trained in simulation using privileged supervision, yet operates without explicit dynamic models or extensive domain randomization. Hardware experiments on a full-scale humanoid validate four key capabilities: Accurate online payload identification using only onboard sensors; improved motion tracking via load-aware control adaptation; robust dual-arm payload transport of up to 20 kg, outperforming existing baselines; and enhanced disturbance rejection via emergent postural strategies. By embedding physical parameter awareness into end-to-end policy learning, this framework offers a scalable and robust solution for real-world humanoid control under dynamic load conditions.
Keywords:
Humanoid control
inertial parameter identification
payload adaptation
reinforcement learning
state estimation
Journal
I
IF:
7.3
Papers:
5.4K
Citations:
2.4W
Organization
Cited Papers