Learning-based Adaptive Control of Quadruped Robots for Active Stabilization on Moving Platforms

TL;DR

This work tackles the problem of maintaining balance for a quadruped on moving platforms with $6$-$\mathrm{DoF}$ motions and unknown platform dynamics. It introduces Learning-based Active Stabilization on Moving Platforms (LAS-MP), featuring a self-balancing policy and two state estimators trained via Regularized Online Adaptation, complemented by platform-velocity alignment commands and B-spline-based trajectory generation with curriculum learning. The approach is validated in simulation against multiple baselines, with ablations demonstrating the value of explicit state estimation and the alignment command, and the estimators showing accurate state and intrinsic-property inference. The results indicate that LAS-MP achieves superior balancing performance across diverse platform motions, highlighting its potential for real-world deployment and prompting further real-world validation and exteroceptive-sensor integration to address steady-state deviations.

Abstract

A quadruped robot faces balancing challenges on a six-degrees-of-freedom moving platform, like subways, buses, airplanes, and yachts, due to independent platform motions and resultant diverse inertia forces on the robot. To alleviate these challenges, we present the Learning-based Active Stabilization on Moving Platforms (\textit{LAS-MP}), featuring a self-balancing policy and system state estimators. The policy adaptively adjusts the robot's posture in response to the platform's motion. The estimators infer robot and platform states based on proprioceptive sensor data. For a systematic training scheme across various platform motions, we introduce platform trajectory generation and scheduling methods. Our evaluation demonstrates superior balancing performance across multiple metrics compared to three baselines. Furthermore, we conduct a detailed analysis of the \textit{LAS-MP}, including ablation studies and evaluation of the estimators, to validate the effectiveness of each component.

Figures (7)

• Figure 1: An illustration demonstrating a balancing process of a quadruped robot on a moving platform exhibiting six-degrees-of-freedom motions.
• Figure 2: Overall framework of the Learning-based Active Stabilization method on Moving Platforms (LAS-MP). It consists of four key components: (1) parallelized simulation environments for moving platforms, (2) a platform trajectory generator with a scheduling mechanism for managing task complexity progression, (3) a reinforcement learning (RL) algorithm for policy optimization, and (4) a self-balancing policy with two state estimators. To leverage the privileged information in policy learning, we concurrently train the self-balancing policy with the two system state estimators in a single phase by employing the Regularized Online Adaptation (ROA) method cheng2023parkour. To clearly differentiate components utilized in training or evaluation phases, we designate yellow for training components and purple for evaluation components. A combination of both colors represents components involved in both phases.
• Figure 3: This figure visually presents the notations used in the descriptions, such as the coordinate frames for the body $\mathcal{B}$, platform $\mathcal{P}$, and world $\mathcal{W}$. It also illustrates estimated robot and platform states, including contact states $\hat{\bm{c}}_{ee}$ and linear velocities of the body $\hat{\bm{v}}^{\mathcal{B}}_{body}$ and platform $\hat{\bm{v}}^{\mathcal{B}}_{plf}$. Platform velocities are visualized after conversion to the platform frame for clarity.
• Figure 4: Examples of training platform trajectories $\Xi_{\text{train}}$ in translational space and their statistical analysis. A similar trend is observed in rotational space.
• Figure 5: Snapshots undergoing an evaluation process of the LAS-MP, captured with a fixed camera in the world coordinate frame $\mathcal{W}$.