Bytedance presented a shi brain for household robots

BYTEDANCE, which owns TikTok, presented the GR-3 system-a multimodal SI-model of the type “stretch-speech”, which acts as a brain for robots. She allows them Understand human teams and perform daily actions – for example, to clean a table or hang clothes.

GR-3 provides robots with the ability to work with unfamiliar objects, navigate the new environment and respond to abstract concepts-such as size or spatial location. The system makes it possible to perform tasks according to a single language instruction, without the need for preliminary training in specific subjects.

In a demonstration video, a two -hand Bytemini Bytemini robot inserts a hanger into a shirt and gently places it on a rack.

🚀🚀🚀 Ever Wonderned What It Takes for Robots to Handle Real-World Household Tasks? Long-Horizon Execution, Deformable Object Dexterity, and Unseen Object Generalization-Meet GR-3, Bytedance Seed’s New Vision-Language-Ection (VLA) Model!

GR-3 IS A GENAralizable… pic.twitter.com/zecrjaxc0j

– Xiao ma (@yusufma555) July 22, 2025

In the technical report, the team noted that the robot successfully works with short sleeves, although the training data had only long things. The system can find objects not only by the name, but also by characteristics such as “Big Player” or “One on the left”.

BYTEDANCE has used a quantity approach to GR-3 training, which includes:

• joint training on large -scale images and text dates;
• Dophineing in human trajectories collected through VR devices;
• Setting up the model by simulating Android actions.

“We hope that GR-3 will be a step towards creating universal robots that can help people in everyday life,” the developers said.

Recall that in January, Perplexity AI startup suggested by BYTEDANCE combine the American Tiktok, Perplexity itself and new investors into a common legal structure.

The gun