[論文レビュー] Learning Transferability: A Two-Stage Reinforcement Learning Approach for Enhancing Quadruped Robots' Performance in U-Shaped Stair Climbing

Baixiao Huang, Baiyu Huang+1 more|arXiv (Cornell University)|Feb 16, 2026

Robotic Locomotion and Control被引用数 0

ひとこと要約

The paper presents a two-stage end-to-end deep reinforcement learning approach where a quadruped robot is trained on pyramid-stair terrain and then applied to U-shaped stairs, demonstrating policy transferability across stair geometries and models.

ABSTRACT

Quadruped robots are employed in various scenarios in building construction. However, autonomous stair climbing across different indoor staircases remains a major challenge for robot dogs to complete building construction tasks. In this project, we employed a two-stage end-to-end deep reinforcement learning (RL) approach to optimize a robot's performance on U-shaped stairs. The training robot-dog modality, Unitree Go2, was first trained to climb stairs on Isaac Lab's pyramid-stair terrain, and then to climb a U-shaped indoor staircase using the learned policies. This project explores end-to-end RL methods that enable robot dogs to autonomously climb stairs. The results showed (1) the successful goal reached for robot dogs climbing U-shaped stairs with a stall penalty, and (2) the transferability from the policy trained on U-shaped stairs to deployment on straight, L-shaped, and spiral stair terrains, and transferability from other stair models to deployment on U-shaped terrain.

研究の動機と目的

• Motivate autonomous stair climbing for quadruped robots in building construction tasks.
• Develop an end-to-end RL framework to optimize performance on U-shaped stairs.
• Explore transferability of learned policies across different stair geometries and models.

提案手法

• Propose a two-stage end-to-end deep RL approach for a quadruped robot (Unitree Go2).
• Stage 1: train on pyramid-stair terrain in Isaac Lab to learn climbing policies.
• Stage 2: transfer the learned policies to climbing a U-shaped indoor staircase.
• Evaluate policy effectiveness with stall penalties to reach climbing goals.
• Demonstrate cross-geometry and cross-model transferability (straight, L-shaped, spiral stairs; other stair models).

実験結果

リサーチクエスチョン

• RQ1 Can an RL policy trained on pyramid-stair terrain transfer effectively to U-shaped indoor stairs?
• RQ2 Does the learned policy exhibit transferability to straight, L-shaped, and spiral stair terrains?
• RQ3 Does transferability extend across different stair models beyond the training environment?

主な発見

• The approach enables a successful goal achievement for robot dogs climbing U-shaped stairs with stall penalties.
• Policies trained on U-shaped stairs transfer to straight, L-shaped, and spiral stair terrains.
• Policies trained on other stair models transfer to U-shaped terrain.
• The work demonstrates the viability of end-to-end RL methods for autonomous stair climbing in quadruped robots.