GR-1
- joint prediction
- 2023
- GPT-style video pretraining
- World action model
GR-1 is a GPT-style transformer that first learns to predict video, then is fine-tuned to emit robot actions and future frames together.
Key facts
Rows below are the register record in lib/models.ts. A dash means the value is not confirmed from a primary source. Classification follows Reuss, NVIDIA Technical Blog (15 June 2026) and the model's own paper.
| Name | GR-1 |
|---|---|
| Lab | — |
| Year | 2023 |
| Paradigm | joint prediction |
| Backbone | GPT-style video pretraining |
| Source | https://arxiv.org/abs/2312.13139 |
What the paper claims
- GR-1 takes a language instruction, a sequence of observation images and robot states, and predicts robot actions as well as future images in one model. Wu et al., 2023
- It is first pretrained on video prediction with a large-scale video dataset, then fine-tuned on robot data without changing the architecture. Wu et al., 2023
- Reuss treats GR-1 as the early joint-prediction WAM: a GPT-style policy with video-prediction readout tokens, later scaled up by models such as DreamZero. Reuss, NVIDIA Technical Blog (15 June 2026)
What it evaluates
- On the CALVIN benchmark the authors report raising success rate from 88.9% to 94.9% against the previous methods in their comparison, and average completed-task length in a row of five from 3.06 to 4.21. Wu et al., 2023
- In their zero-shot unseen-scene setting they report raising success rate from 53.3% to 85.4%. With 10% of the training data they report 77.8% success against 66.8% for the best baseline in that table. Wu et al., 2023
- They also report real-robot experiments on object transport and articulated-object manipulation, including tests on unseen scenes and objects. Wu et al., 2023
- Reuss highlights the harder CALVIN ABC→D split in the GR-1 table: prior methods stayed below an average sequence length of 1.0, while GR-1 reached 3.06 of 5. Reuss, NVIDIA Technical Blog (15 June 2026)
On this register
Joint prediction. Emits video and actions together. That axis is defined in what is a world action model? and in Reuss, NVIDIA Technical Blog (15 June 2026).
Other joint prediction models on the register: GENIMA, Unified Video Action, Cosmos Policy, DreamZero.
No other register row currently lists the backbone GPT-style video pretraining. Matching is exact; nearby names such as Wan 2.1 versus Wan 2.2 are kept separate.