Reproduce “Sample-Efficient Robot Skill Learning for Construction Tasks: Benchmarking Hierarchical Reinforcement Learning and Vision-Language-Action VLA Model” on OpenResearch

About this paper

Zhaofeng Hu, Hongrui Yu, Vaidhyanathan Chandramouli, Ci-Jyun Liang

Researchers at Stony Brook University and Virginia Tech evaluated Hierarchical Reinforcement Learning (HRL) and Vision-Language-Action (VLA) models for multi-stage construction tasks, finding VLA models offer superior generalization and few-shot learning for rapidly adapting to new scenarios with reduced programming. This comparison provides a cost-benefit analysis for automating dexterous material manipulation and installation in construction.

Key insights

• Vision-Language-Action models, particularly π₀.₅, leverage large-scale pre-training to offer strong few-shot generalization and robustness for variable construction tasks, significantly reducing new programming effort.
• Hierarchical Reinforcement Learning models (DQN) can achieve high performance with less initial data but often require additional tuning, such as noise injection, to ensure robust behavior.
• Architectural features like multi-view perception and continuous action heads are critical for precise and smooth robot control in complex manipulation tasks, as evidenced by OpenVLA's failure due to their absence.

Summary by alphaXiv.