VLARLKit

An elegant Reinforcement Learning library for Vision-Language-Action models.

VLARLKit is a research toolkit for applying Reinforcement Learning to Vision-Language-Action (VLA) models. We focus on clean, efficient, and reproducible implementations for robotic manipulation research.

The core training framework. Highlights:

• Off-policy async training with daemon rollout threads and bounded queues
• Significant rollout efficiency gains vs. existing VLA-RL pipelines on LIBERO
• Native support for action chunking and modern VLA architectures
• Compatible with LIBERO, ManiSkill, and other manipulation benchmarks

Maintained by Yihao Sun — yihao.sun@mila.quebec