# Data for embodied intelligence

The complete package for robotics foundation models. Video, trajectories, and rich multimodal annotations — across every configuration, from pre-training to post-training to evals.

## Full-stack robotics data

We cover every stage of the robotics ML pipeline.

### Pre-training

Diverse environments, objects, and manipulation tasks at scale for building foundational robotics models.

### Post-training

Expert teleoperation with precise action labels optimized for imitation learning.

### Evals

Standardized tasks across manipulation and navigation for measuring real-world performance.

## Data across configurations

Multiple camera perspectives and collection modalities.

### Ego-centric

First-person camera mounted on the robot head or body. Captures the robot's perspective during task execution.

### Overhead

External cameras positioned above the workspace. Bird's-eye view of the full scene and robot motion.

### Ego-centric + Wrist

Combined head and wrist-mounted cameras. Full context plus close-up manipulation view for detailed grasping tasks.

### Teleoperation

Human-controlled demonstrations with expert operators. High-quality trajectories optimized for imitation learning.

## AI-powered diversity steering

Our robotics data engine automatically categorizes, validates, and steers collection to continuously improve diversity across every dimension that matters.

### Ingest

Raw data streams in from collection sites worldwide

### Categorize

AI auto-tags environment, objects, tasks, and operator style

### QA + Validate

Automated quality checks and human review for edge cases

### Steer

System identifies gaps and redirects collection to fill them

### Environment coverage

Kitchens, offices, warehouses, retail spaces, and homes. Captured across varied lighting, clutter, and indoor–outdoor conditions.

### Object diversity

Thousands of unique instances across materials (rigid, deformable, transparent), sizes, shapes, and weights.

### Task variety

Pick and place, stacking, insertion, tool use, navigation, and multi-step sequential tasks.

### Operator diversity

Multiple demonstrators per task with varied skill levels, styles, handedness, and demonstration speeds.

## Robotics sample data

Explore sample robotics footage captured across tasks, sensors, and environments.

### Human egocentric: Mono

Washing dishes

### Human egocentric: Stereo

Groceries

### Human egocentric: Trio

Folding clothes

### Bimanual stationary

Stacking plates

### Human egocentric: Trio + Gripper

Packing garments

## NEXT-GEN ROBOTICS

During his visit to our SF robotics lab, Dwarkesh got a firsthand look at how we’re redefining robotics development with cutting-edge data collection.