Skip to content

GR-2

A general-purpose robot agent for manipulation

Visit official site

What GR-2 is

The model is a general-purpose robot agent pretrained on internet video to capture world dynamics, then fine-tuned on robot trajectories.

What you can do with it

  • Perform end-to-end binary picking in industrial settings
  • Work on long-horizon language control in the CALVIN benchmark
  • Operate in new backgrounds, environments, objects and tasks

Model notes

  • Pretraining covers 38 million video clips and 50 billion tokens
  • Fine-tuning targets video generation and action prediction
  • The agent scales efficiently as model size increases

Pros & cons

✓ What we like

  • Generalizes to unseen backgrounds, objects and tasks
  • Multitask learning spans more than 100 tasks

! What to watch out for

  • Pretraining relies on internet video at large scale
  • Fine-tuning needs robot trajectory data

FAQ

How large is the pretraining data?

The material lists 38 million video clips and more than 50 billion tokens.

What success rate is reported?

A 97.7% average success rate across more than 100 tasks is listed.

Last reviewed: 2026-11-15

More LLM API platform tools

View all →

How we review