Glossary
What Is GR-3?
Last reviewed: October 2026
Short answer
GR-3 is a large-scale vision-language-action (VLA) model for generalist robot policies from ByteDance, described in the GR-3 Technical Report (arXiv 2507.15493). It is trained by co-training with web-scale vision-language data, fine-tuning on human trajectories collected with VR devices, and imitation learning on robot trajectories. Firsthand is not part of it.
01
What is GR-3?
GR-3 is introduced in the GR-3 Technical Report (arXiv 2507.15493), first submitted on 21 July 2025, with an official ByteDance Seed project page dated 22 July 2025. The abstract describes a large-scale VLA model that generalizes to novel objects, environments and instructions involving abstract concepts, and can be fine-tuned with minimal human trajectory data. Read: the arXiv abstract page and the project page only. The full paper is not summarised here.
02
Where does Firsthand fit next to GR-3?
GR-3 is third-party research. Firsthand is not affiliated with its authors and claims no use, training or compatibility with it. Firsthand captures first-person human video, not robot data, fast, on demand and to a written spec: 21 keypoints per hand with a visibility flag, wrist trajectories and action segments, plus one task instruction per episode.
03
How is GR-3 trained?
The abstract names a three-part recipe: co-training with web-scale vision-language data, efficient fine-tuning from human trajectory data collected via VR devices, and imitation learning with robot trajectory data. The project page adds that the human trajectories were constructed by lab members using a VR device. Architecture details, parameter counts and dataset sizes are Not stated in the sources read.
04
What does the project page say GR-3 was tested on?
The page reports three challenging real-world tasks on the ByteMini robot, a bi-manual mobile robot introduced in the report. Pick-and-place is tested in four settings: Basic, Unseen Environments (four environments), Unseen Instructions (abstract concepts) and Unseen Objects (45 objects). The long-horizon task is cleaning up a dining table, prompted either with one general instruction or with sub-task instructions in a row. The deformable task is hanging clothes on a drying rack. The abstract states GR-3 surpasses the π0 baseline on a wide variety of tasks. Scores are the authors’ own and are not reproduced here.
05
How do human trajectories help GR-3 adapt to new objects?
The project page reports that GR-3 adapts to objects unseen in the robot trajectory data with few-shot human trajectories. Performance on unseen objects keeps improving from one to ten human trajectories per object, with no significant impact on seen objects. It also reports a significant drop on Unseen Instructions and Unseen Objects when co-training is excluded. See the site entries for co-training and human demonstration data for the general terms.
06
How does GR-3 differ from GR-1 and GR-2?
The GR-1 and GR-2 entries describe models pre-trained on video that predict video and actions together. The GR-3 abstract and project page describe a VLA model co-trained with web-scale vision-language data and do not mention video generation. Whether GR-3 builds on GR-1 or GR-2 is Not stated in the sources read. Firsthand ranks none.
07
What do the GR-3 sources not state?
Parameter counts, architecture, pre-training dataset sizes, hours of robot or human trajectories, action space, weights and licence are Not stated in the sources read. No GitHub repository is linked from the arXiv page or the project page read. Whether the human trajectories are egocentric is Not described.
08
What human data supports VLA fine-tuning, and how do you get it custom-collected?
GR-3’s authors collected human trajectories with VR devices. Firsthand captures human first-person video, not robot data, and does not claim its capture substitutes for VR-collected trajectories. Write tasks, environments and signals into a spec and confirm them on a Spec Pilot. Median time to first validated batch: 48 hours where coverage is deep, three to four weeks where thin. Formats: RLDS, LeRobot v2.1, WebDataset, HDF5; mapping to GR-3’s layout is the buyer’s pipeline.
09
GR-3 components and training recipe, as stated in the official sources
| Item | What the official sources state |
|---|---|
| Model type | Large-scale vision-language-action (VLA) model (abstract) |
| Training, part 1 | Co-training with web-scale vision-language data |
| Training, part 2 | Efficient fine-tuning from human trajectory data collected via VR devices |
| Training, part 3 | Imitation learning with robot trajectory data |
| Robot | ByteMini, a bi-manual mobile robot introduced in the report |
| Evaluation tasks | Three on the project page: pick-and-place, long-horizon table cleanup, hanging clothes |
| Baseline named | π0 (abstract) |
| Architecture, parameters, data sizes, weights | Not stated in the sources read |
| Firsthand relationship | None |
11
GR-3 beside GR-2 and GR-1, as each official source or site entry states
| Item | GR-3 (sources read) | GR-2 (site entry) | GR-1 (site entry) |
|---|---|---|---|
| Described as | Large-scale VLA model | Generative video-language-action model | GPT-style model for language-conditioned manipulation |
| Pre-training data | Web-scale vision-language data in co-training; size Not stated | Daily human activity video; 38 million clips | Ego4D clips; 800,000 clips |
| Human data | Trajectories collected via VR devices | Not described | Ego4D video in pre-training |
| Robot data | Robot trajectories; ByteMini robot | Robot trajectories; size Not stated | CALVIN and a real robot |
| Generates video | Not described in the sources read | Yes, per the project page | Predicts future images |
12
What GR-3 consumes beside what the Firsthand site states
| Aspect | GR-3 (sources read) | What the Firsthand site states |
|---|---|---|
| Human data | Trajectories collected via VR devices | Head-mounted egocentric video plus two wrist cameras; suitability for GR-3 not claimed |
| Actions | Robot trajectories on ByteMini | Robot actions Not stated; 21 keypoints per hand, visibility flag, wrist trajectories, action segments |
| Language | Instructions, including abstract concepts; text format Not stated | One task instruction per episode |
| File layout | Not stated in the sources read | RLDS, LeRobot v2.1, WebDataset, HDF5 |
| Compatibility | Not claimed | Not claimed |
13
Frequently asked questions
- What is GR-3?
- A large-scale vision-language-action model for generalist robot policies, described in the GR-3 Technical Report (arXiv 2507.15493).
- How is GR-3 trained?
- By co-training with web-scale vision-language data, fine-tuning on human trajectories collected via VR devices, and imitation learning on robot trajectories.
- What robot does GR-3 run on?
- The report introduces ByteMini, a bi-manual mobile robot, and the abstract says it can accomplish a wide range of tasks when integrated with GR-3.
- Is GR-3 open source?
- The arXiv page and project page read link no code repository or weights. Availability and licence are Not stated in the sources read.
- Was GR-3 trained on egocentric video?
- Not described in the sources read. The human trajectories were collected via VR devices.
- Does Firsthand use GR-3, and can I request a custom dataset?
- Firsthand is not affiliated with the GR-3 authors and claims no use, training or compatibility. You can request a custom dataset, confirmed first on a Spec Pilot.