ByteDance Seed Team Unveils GR-RL, Pushing VLA Models Toward Long-Horizon Dexterous Manipulation
Want to read in a language you're more familiar with?
ByteDance’s new GR-RL framework marks a major leap in robot dexterity, achieving the first continuous robot shoelace-threading demonstration with an 83% success rate.
The Seed team at ByteDance has released its latest research breakthrough, GR-RL, aimed at expanding the limits of Vision-Language-Action (VLA) models in long-duration, fine-grained robotic manipulation. GR-RL introduces a reinforcement learning framework that spans offline data filtering to online real-world fine-tuning, and has achieved industry first: enabling a robot to thread an entire shoelace through an entire shoe in one continuous sequence.
Compared with the previous supervised learning model GR-3, GR-RL boosts the success rate of the shoelace-threading task from 45.7% to 83.3%, reducing failures by nearly 70%.
