New DRIVE method enhances VLA policy generalization through diverse success trajectories

Researchers have developed a new method called DRIVE (Diversity-driven RL fIne-tuning for VLA gEneralization) to improve the generalization capabilities of vision-language-action (VLA) policies. This technique uses reinforcement learning to explicitly encourage diversity among successful trajectories, rather than just optimizing for success itself. Experiments on benchmarks like LIBERO-Plus, ManiSkill3, and RoboTwin 2.0 showed DRIVE improved out-of-domain performance by 5.3 points on average. When tested on a physical robot platform, AgileX PiPER-X, DRIVE increased the success rate from 64.1% to 73.3%. AI

IMPACT Enhances AI policy generalization, potentially leading to more robust and adaptable robotic systems in real-world applications.

RANK_REASON The cluster contains a research paper detailing a new method for improving AI policy generalization. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• AgileX PiPER-X
• arXiv
• DRIVE
• LIBERO-Plus
• ManiSkill3
• reinforcement learning
• RoboTwin 2.0
• Vision-language-action policies

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →