A Complete Data Development Solution for Physical AI

A Complete Data Development Solution for Physical AI

Reliable and scalable data collection is a bottleneck for the development of Large World Models and Robot Foundation Models. By ensuring high-quality gold standard datasets and tools optimized for internet-scale data collection capabilities, Dexter offers a comprehensive solution to accelerate Physical AI.


Think of Dexter as an elite data concierge, meticulously curating high-fidelity world model and robotics training data—tailored for any modality, task, or scale with precision, speed and sophistication

Reliable and scalable data collection is a bottleneck for the development of Large World Models and Robot Foundation Models. By ensuring high-quality gold standard datasets and tools optimized for internet-scale data collection capabilities, Dexter offers a comprehensive solution to accelerate Physical AI.


Think of Dexter as an elite data concierge, meticulously curating high-fidelity world model and robotics training data—tailored for any modality, task, or scale with precision, speed and sophistication

How ?

How ?

To solve the data bottleneck problem, we are building at the intersection of human skills & sophisticated tooling.


Our tool-enabled service accelerates Embodied AI development and World Model creation by unifying high-quality data collection across teleoperation, simulation, video, and motion capture, with seamless API access. To scale operations, we maintain a network of vetted professionals for specialized task data collection, remote operators to streamline teleoperation, and 3D modelers for diverse robotics applications.


Our goal? Very high-quality ready to use labeled training data for any task significantly reducing both time and cost.

To solve the data bottleneck problem, we are building at the intersection of human skills & sophisticated tooling.


Our tool-enabled service accelerates Embodied AI development and World Model creation by unifying high-quality data collection across teleoperation, simulation, video, and motion capture, with seamless API access. To scale operations, we maintain a network of vetted professionals for specialized task data collection, remote operators to streamline teleoperation, and 3D modelers for diverse robotics applications.


Our goal? Very high-quality ready to use labeled training data for any task significantly reducing both time and cost.

Simulation

Dxtr Tool

dxtr.app

Dxtr Pro Service

dxtr.pro

Simulation

Ability to choose from popular simulation engines

Unified data formats across modalities

Quick & easy scene, episode generation

Expert 3d modelers to create desired assets and scene authoring

Video2Embodiment

Integration with models that offers different levels of autonomy to go from human ego centric videos to targeted robot hardware

Tool that is optimized for relevant data collection and supports both mobile and stationary cameras 

Gives the ability to achieve internet scale data generation

Access to specialists around the world in multitude of professions (e.g: Carpenters, plumbers etc)

Quick turn around with large network of data collectors

Teleoperation

Supported for widely used cobots and standard robot platforms

Expert Pilots trained in teleoperation available on demand

Motion Capture

Currently not supported

Currently not supported

Representative

Robot HW

Currently not supported

Currently not supported

Simulation

Video2Embodiment

Teleoperation

Motion Capture

Currently not supported

Representative

Robot HW

Currently not supported

Video2Embodiment Models

Video2Embodiment Models

Our solution enables video-driven data collection with rich, multi-layered metadata, allowing seamless retargeting of real-world human actions to robots while minimizing the embodiment gap.

Our solution enables video-driven data collection with rich, multi-layered metadata, allowing seamless retargeting of real-world human actions to robots while minimizing the embodiment gap.

Original Video

Original Video — Egocentric video, produced using mobile phone without any specialized setup.

Original Video — Egocentric video, produced using mobile phone without any specialized setup.

Spatial & semantic awareness

Spatial & semantic awareness

Dexter enhances spatial cognition, generating 3D maps from videos to extract metric depth and improve understanding of surroundings and robot planning for manipulators, bimanual robots, bipeds, and mobile platforms. Semantic awareness further refines understanding of tasks and outcomes.

Dexter enhances spatial cognition, generating 3D maps from videos to extract metric depth and improve understanding of surroundings and robot planning for manipulators, bimanual robots, bipeds, and mobile platforms. Semantic awareness further refines understanding of tasks and outcomes.

Scene understanding

Inbuilt segmentation & object labeling for a richer metadata informing scene understanding. Optimized tooling for faster data curation and QA with humans in the loop.

Inbuilt segmentation & object labeling for a richer metadata informing scene understanding. Optimized tooling for faster data curation and QA with humans in the loop.

3d map creator

Environmental reconstruction for enhanced spatial understanding, enabling richer metadata and reduced embodiment gap in task planning and execution

Environmental reconstruction for enhanced spatial understanding, enabling richer metadata and reduced embodiment gap in task planning and execution

Action chunking and detailed annotations

Action chunking and detailed annotations

Optimized data collection workflows deliver highly detailed, structured annotations in a semi-automated process, whether sourced from Video2Embodiment, simulation, or other methods.

Optimized data collection workflows deliver highly detailed, structured annotations in a semi-automated process, whether sourced from Video2Embodiment, simulation, or other methods.

00:00

00:00

Approaching Stove

Approaching Stove

00:04

00:04

Finding and reaching target knob

Finding and reaching target knob

00:06

00:06

Turning the knob counter clockwise by ~90 degrees

Turning the knob counter clockwise by ~90 degrees

00:08

00:08

Taking hand back

Taking hand back

Seamless integration into ML pipelines

Seamless integration into ML pipelines

A simple, flexible API lets you manage data, request datasets from our workforce, define ontologies for annotation and integrate internal processes with ease.

A simple, flexible API lets you manage data, request datasets from our workforce, define ontologies for annotation and integrate internal processes with ease.

Workflow 1

Workflow 2

Workflow 3

+

1

2

3

4

5

6

7

8

9

from dxtr import AnnotationClient


# Initialize API client with user token and project ID

client = AnnotationClient(api_token="your_api_token_here", project_id="project_12345")


# Get status of data collection for a specific project

status = client.get_status(project_id="project_12345")


print(status)

REQUEST Data Collection

POST

Check Collection Status

GET

DEFINE ONTOLOGIES

PUT

Automated Labeling

POST

UPLOAD DATA FOR REVIEW

POST

List Available Datasets

GET

Export Labeled Dataset

GET

EXPORT RAW DATA

POST

View Processing Tasks

GET

Delete Dataset

DEL