iPhone + head mount
· wide ego view
· guided capture in-app
[MISSION] HUMAN DEMONSTRATION DATA FOR PHYSICAL AI
Teach robots_
by just living your life.
We pay people to record first-person video, hand motion, and proprioception data from wearables and phones today — and democratize access to robot teleoperations tomorrow. That data trains the foundation models that move the world's robots.
• A.1 · sort_items_into_bins pick_place• A.1 · stack_objects pick_place• A.1 · organize_shelf pick_place• A.1 · load_dishwasher pick_place• A.2 · press_elevator_button interface• A.2 · operate_microwave interface• A.2 · use_remote_control interface• A.2 · adjust_thermostat interface• A.3 · open_door_with_handle entry_access• A.3 · use_keys_and_locks entry_access• A.3 · slide_door entry_access• A.3 · operate_latch entry_access• A.4 · use_spatula tool_use• A.4 · sweep_with_broom tool_use• A.4 · operate_screwdriver tool_use• A.4 · pour_from_container tool_use• A.5 · fold_towels deformable• A.5 · handle_produce deformable• A.5 · manage_cables deformable• A.5 · wrap_items deformable• A.1 · sort_items_into_bins pick_place• A.1 · stack_objects pick_place• A.1 · organize_shelf pick_place• A.1 · load_dishwasher pick_place• A.2 · press_elevator_button interface• A.2 · operate_microwave interface• A.2 · use_remote_control interface• A.2 · adjust_thermostat interface• A.3 · open_door_with_handle entry_access• A.3 · use_keys_and_locks entry_access• A.3 · slide_door entry_access• A.3 · operate_latch entry_access• A.4 · use_spatula tool_use• A.4 · sweep_with_broom tool_use• A.4 · operate_screwdriver tool_use• A.4 · pour_from_container tool_use• A.5 · fold_towels deformable• A.5 · handle_produce deformable• A.5 · manage_cables deformable• A.5 · wrap_items deformable
>1000
HOURS OF DEMONSTRATION DATA
>500
ACTIVE CONTRIBUTORS
20
TASK CATEGORIES
// 02_HOW_IT_WORKS
Three steps to train the next generation.
Pick a task from the taxonomy, film yourself performing it, earn rewards when your demonstration is accepted.
- 01
Pick a Task
Browse the Deviant task taxonomy in the app — from Pick & Place to Deformable Objects — and choose one that fits the moment you're in.
A.1 – A.5taxonomyeveryday_tasks - 02
Record Yourself
Use the Deviant app to film yourself performing the task. Our guided capture ensures consistent, high-quality training data. Record first-person — from your own point of view, hands in frame. Third-person clips (someone filming you) and clips filmed one-handed while holding the phone earn reduced points — mount your phone so both hands stay on the task.
iPhoneguided_capturefirst_person - 03
Earn Rewards
Accepted demonstrations become part of the Deviant training set. You get rewarded when your clip passes verification.
per_clipverified
// 03_HARDWARE
Works with what you already have.
We're starting with the device in your pocket. More form factors in the pipeline.
Apple Vision Pro
· stereo ego · eye gaze
· LIDAR · scene mesh
· persona_replace privacy
Ray-Ban Meta / Aria
· mono ego · audio beamform
· 12MP @ 30fps
· best for mobile tasks
Manus IMU gloves (add-on)
· 22 joints per hand
· 120Hz · pairs with any ego_cam
· unlocks dexterous capture
Unitree G1 Teleop Rig
· 23-DOF humanoid · bimanual capture
· suspended gantry · safe teleop sessions
· action-labeled trajectory data
// GET_NOTIFIED
Be first in line.
New devices are being added as the app rolls out. Join the waitlist to hear when yours is supported.
join_waitlist →// 04_TASK_TAXONOMY
Real tasks. Real complexity.
Our structured taxonomy covers the full spectrum of manipulation skills robots need to master.
A.1
Pick & Place
Grasping objects and placing them precisely — from sorting groceries to organizing shelves. The foundational skill for robotic manipulation.
EXAMPLE TASKS
—Sorting items into bins
—Stacking objects
—Organizing a shelf
—Loading a dishwasher
A.2
Interface Manipulation
Operating buttons, switches, touchscreens, and controls. Teaching robots to interact with the interfaces humans designed for humans.
EXAMPLE TASKS
—Pressing elevator buttons
—Operating a microwave
—Using a remote control
—Adjusting a thermostat
A.3
Entry & Access
Navigating doors, gates, locks, and physical barriers. Essential for robots that need to move through real-world environments.
EXAMPLE TASKS
—Opening doors with handles
—Using keys and locks
—Sliding doors and gates
—Operating latches
A.4
Tool Use
Wielding handheld tools effectively — from kitchen utensils to cleaning equipment. Complex dexterous manipulation in action.
EXAMPLE TASKS
—Using a spatula
—Sweeping with a broom
—Operating a screwdriver
—Pouring from a container
A.5
Deformable Objects
Handling soft, flexible, and deformable materials — one of the hardest challenges in robotics. Fabrics, food, cables, and more.
EXAMPLE TASKS
—Folding towels and clothes
—Handling produce
—Managing cables and wires
—Wrapping items
// 05_REWARDS
Get paid for the data you generate.
Higher quality and longer videos result in more rewards.
// 06_ROADMAP
Where this is going.
Egocentric capture is Phase 1. The network is built to expand across the full lifecycle of robot foundation model training — and the value of holding a piece of it grows with each stage.
EGO_CAM · A.2_wipe_clean
Egocentric Capture
iPhone + head mount. First-person video of household tasks. The modality used to pretrain embodied AI world models — the foundation every robotics foundation model is built on.
TELEOP · arm_so100
Teleoperation
Remote-control simulated and physical robot arms. Capture action-labeled trajectory data — denser, more structured, and significantly more valuable per minute than passive video. The modality used to fine-tune a pretrained world model into a robot that performs specific tasks reliably.
// classified
█████████████
██████ ████
[ to be announced ]
—
Together, Phases 1 and 2 give Deviant coverage across the entire training pipeline — pretraining and fine-tuning. The dataset becomes the primary input for both stages of robot foundation model development. Holders own a share of an asset whose commercial relevance compounds with every layer the network adds.
// 07_FAQ
Common questions.
What hardware do I need?
Just an iPhone and a head mount. Deviant is designed so you can pull out your phone, pick a task, and record. Refer to our whitepaper for examples of head mounts you can use and best practices for recording.
How does earning work?
You earn rewards for demonstrations that pass verification — the higher the quality, the larger the rewards. Exact amounts, tiers, and payout methods are subject to change.
Where can I read more?
What will the recordings be used for?
Data collected is used to train cutting edge foundation models with partners like research groups, robotics labs, industry leaders and more.
What kind of tasks can I record?
There are currently five task categories: Pick & Place, Interface Manipulation, Entry & Access, Tool Use, and Deformable Objects. Each category covers common everyday skills — see the tasks section above for examples. More tasks will be added on an ongoing basis based on demand from our partners.