GR-2
Generative video-language-action model pre-trained on 38M internet videos, then fine-tuned for robot manipulation.
About
GR-2 comes from ByteDance's Seed robotics team and takes an unusual route to robot control: it starts as a video generator. The model is first pre-trained on 38 million internet video clips (over 50 billion tokens) to learn how the world moves, then fine-tuned jointly for video generation and action prediction on robot trajectories. The bet is that predicting the next frames of a task and predicting the actions that cause them are two sides of the same coin. The authors report 97.7% average success across more than 100 manipulation tasks and strong generalization to new backgrounds, objects, and environments. All of that is their own measurement. The project page is live, but I found no public code or weight release for GR-2 as of this writing (GR-1, its predecessor, was open-sourced; GR-3 followed in 2025). Paper: arXiv 2410.06158, October 2024.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Paper (arXiv 2410.06158): 97.7% avg success across 100+ tasks (author-reported)
- No independent benchmarks and no public weights to verify against
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (TIER metrics, metrics-v1). Evidence: ByteDance Seed paper (arXiv 2410.06158) with a distinctive video-pretraining recipe and broad task coverage. Strengths: web-scale video knowledge, strong reported results, credible follow-through (GR-3 in 2025). Docked for: (1) no public code or weights, so nothing independently checkable; (2) all benchmarks are author-reported. License: none — no public release.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 6.0/10
- Openness · license + docs
- 4.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · ByteDance Seed paper arXiv 2410.06158 (Oct 2024); pre-trained on 38M internet video clips (50B+ tokens)
- · 97.7% avg success across 100+ manipulation tasks (author-reported)
- · GR-1 was open-sourced; no public code/weights found for GR-2 as of research date