GRTraining data for Physical AI & LLMs
Real-world data for machines that learn.
Graviti captures, annotates and licenses production-grade datasets for humanoids, robots and large language models — with consent and provenance built into every record.
Active capture locations
- Egocentric video
- 3D hand pose
- Teleoperation
- Multi-view RGB-D
- IMU
- Language narration
- Expert SFT
- Preference pairs
- Agent trajectories
01Flagship dataset
Egocentric Household Activities. First-person video of real people doing real chores in real homes.
Head-mounted video of everyday household tasks — cooking, cleaning, laundry, tidying — captured by consented contributors in their own homes across dozens of countries. Every clip ships with dense action segments, 3D hand pose and natural-language narration, ready for training vision-language-action models and humanoid policies.
- Hours of video
- 10,000+
- Contributors
- 2,400+
- Countries
- 38
- Task taxonomy
- 600+ tasks
02Global capture network
A global effort to record how the world actually works.
We don't stage data in labs. Contributors and partner businesses capture real tasks in real environments across six continents, so your models see the diversity they will meet in deployment.
03Catalog
Two catalogs. One standard of quality.
Off-the-shelf datasets you can license today, each with a public spec and a downloadable sample.
Physical AI
For humanoids, manipulation and world models.
Egocentric Household Activities
First-person video of real people doing real chores in real homes.
- Hours of video
- 10,000+
- Contributors
- 2,400+
Bimanual Teleoperation Trajectories
Synchronized multi-view robot demonstrations with language instructions.
- Episodes
- 120,000+
- Task families
- 180
Egocentric Skilled Work
First-person footage of trained professionals in commercial environments.
- Target hours
- 5,000
- Workplace types
- 14
Large language models
For reasoning, alignment and agents.
Expert Reasoning SFT
Multi-step solutions written and verified by domain experts.
- Examples
- 250,000+
- Domains
- 24
Human Preference Pairs
Pairwise and graded comparisons for RLHF, DPO and reward modeling.
- Comparisons
- 400,000+
- Avg. turns
- 4.2
Agentic Tool-Use Trajectories
Human-demonstrated multi-step tasks across browsers, terminals and APIs.
- Trajectories
- 30,000+
- Environments
- 70+
04How we build data
From the real world to your training run.
One accountable pipeline, from capture hardware to delivered shards. The same pipeline powers our catalog and custom programs.
- 01
Capture
A vetted global network of contributors, operators and domain experts records data on calibrated devices — in homes, workplaces and labs.
Head-mounted cameras · teleop rigs · expert authors
- 02
Annotate
Trained annotators and model-assisted tooling add hand pose, action segments, language, preferences and rationales.
Hand pose · actions · narration · rubrics
- 03
Verify
Multi-pass human review, automated checks and contamination screens. Every batch ships with measured quality metrics.
Two-pass QA · agreement metrics · audits
- 04
Deliver
Clean rights and provenance, delivered in the formats your training stack already reads — to your bucket, on your schedule.
LeRobot · RLDS · WebDataset · Parquet · JSONL
05Delivery
Ready for your stack on day one.
Standard formats, documented schemas and data cards. Delivered to S3, GCS, Azure or Hugging Face — no conversion scripts required.
- LeRobot
- RLDS
- WebDataset
- Parquet
- JSONL
- MP4
from lerobot.datasets import LeRobotDataset
# Point at the extracted Graviti sample
ds = LeRobotDataset("graviti/teleop-manipulation-sample", root="./sample")
frame = ds[0]
frame["observation.images.head"].shape # (3, 480, 640)
frame["action"].shape # (14,)
ds.meta.tasks[0] # 'Place the red cup on the top shelf'06Rights & compliance
Data you can actually ship a model on.
Clean rights are not an afterthought. Read how we license data on our company page.
Consent by design
Every contributor signs a written, revocable agreement that explicitly covers AI training.
Provenance per record
Each clip, episode or example traces back to its capture session, device and consent record.
Privacy protection
Faces, screens and documents are blurred; PII is removed and audio excluded unless licensed.
Clear commercial license
Perpetual rights to train and deploy models, with exclusive options for strategic partners.
07FAQ
Questions we hear most.
What is Graviti?
Graviti is a training data company that builds and licenses production-grade datasets for Physical AI (humanoids, robotics, world models) and large language models, including egocentric video, robot teleoperation trajectories, expert-written SFT data, human preference data and agentic trajectories.
What datasets does Graviti offer?
Graviti's catalog covers two areas. Physical AI: Egocentric Household Activities (available), Bimanual Teleoperation Trajectories (early access) and Egocentric Skilled Work (in collection). LLM: Expert Reasoning SFT (available), Human Preference Pairs (available) and Agentic Tool-Use Trajectories (early access). Every dataset has a public specification at graviti.com/datasets and an evaluation sample.
Does Graviti provide training data for humanoid robots and robotics?
Yes. Graviti's Physical AI datasets include first-person (egocentric) video of real household and workplace tasks with 3D hand pose, action segments and narration, and multi-view robot teleoperation demonstrations with joint states and language instructions. They are used to pre-train vision-language-action (VLA) models and to train manipulation policies for humanoids and robot arms.
Does Graviti provide data for training large language models?
Yes. Graviti offers expert-written supervised fine-tuning (SFT) data with step-by-step reasoning, human preference pairs for RLHF, DPO and reward modeling, and human-demonstrated agentic tool-use trajectories. LLM data is authored and verified by domain experts and checked for contamination against public benchmarks.
Can I get a free sample before licensing a dataset?
Yes. Each dataset page has a sample request form. Samples are free, provided for internal evaluation only, and delivered as a download link after a short request with a work email address.
Get started
Tell us what your model needs to learn.
Share your target tasks, modalities and volume. We reply within one business day with matching datasets, samples and a quote.