SimpleVLA-RL: Scaling VLA with RL скачать в хорошем качестве

SimpleVLA-RL: Scaling VLA with RL

In this AI Research Roundup episode, Alex discusses the paper: 'SimpleVLA-RL: Scaling VLA Training via Reinforcement Learning' This work introduces SimpleVLA-RL, an online RL framework to scale Vision–Language–Action models for robotic manipulation without relying on large human-trajectory SFT datasets. It adapts veRL and GRPO with VLA-specific rollouts, parallel rendering, and a simple binary outcome reward spread across action tokens. To boost exploration and stability, it adds dynamic sampling, an asymmetric GRPO clip range, raised rollout temperature, and removes KL regularization to cut memory costs. The goal is better generalization under distribution shift, especially for long-horizon, compositional tasks. Paper URL: https://arxiv.org/abs/2509.09674 #AI #MachineLearning #DeepLearning #VisionLanguageAction #ReinforcementLearning #Robotics Resources: GitHub: https://github.com/PRIME-RL/SimpleVLA-RL