GR-RL
$ 0
GR-RL is ByteDance Seed’s robot foundation model for precise, long manipulation tasks. It combines learning from demonstrations with real-world practice, and it is the first learning-based model to lace a shoe on its own.
Available on backorder
Brain Score
3
Specifications and details:
| Nationality | China |
|---|---|
| Website | https://seed.bytedance.com/en/ |
| Model type | Foundation Model |
| Manufacturer | ByteDance Seed (ByteDance) |
| Release date | December 2025 |
Description
GR-RL is a robot foundation model from ByteDance Seed, the AI research team of TikTok’s parent company. The team introduced it in December 2025. It builds on GR-3, ByteDance’s general vision-language-action model. However, GR-RL targets a specific weakness of such models: very precise, long tasks. Most robot models learn by copying human demonstrations. ByteDance argues that these demonstrations are often noisy and imperfect, especially for delicate work. For that reason, GR-RL adds reinforcement learning, so the robot keeps improving through its own practice.
To show what the model can do, ByteDance chose a famously tricky task: lacing a shoe. The robot has to thread a soft lace through small holes, one after another. According to ByteDance, GR-RL is the first learning-based model that laces a shoe on its own. In tests, it succeeded about 83% of the time. The team trained it in three steps. First, it filtered out poor demonstrations. Next, it mirrored the remaining examples to create more training data. Finally, the robot practiced on its own and learned from its results.
New! 2026 Humanoid
Robot Market Report
198 pages of exclusive insight from global robotics experts — uncover funding trends, technology challenges, leading manufacturers, supply chain shifts, and surveys and forecasts on future humanoid applications.
now Google DeepMind
Contact Humanoid.guide
Website: https://seed.bytedance.com/en/gr_rl
