hand_R · grasp
01 / 05
Mechanic
Repairing an engine
- 01locate valve cover
- 02grasp rocker arm
- 03seat & align
- 04verify clearance
ElectricAI LabsThe data layer for physical AI
The world's experts are already generating the data robots need.
We turn real-world, first-person video of experts at work into structured, robot-ready training data — hands, tools, actions and 3D trajectories.
- hours of egocentric video processed in pilot
- 1,200+
- skill primitives in our taxonomy
- 40+
- annotation layers per frame
- 9
electric-ai / pipeline / session_0412 / cam_ego_01.mp4v0.3
wrench_19mm · 0.97
hand_R · power_grasp
REC00:16.8
reach
grasp
align
torque
verify
structured_output.json
taskreplace_water_pump
subtaskloosen_mount_bolt (3/7)
skilltorque
contactwrench ↔ bolt_M10
grippower · 37 mm
est_torque24.8 N·m
hand_pose21 kp × 2 · 3D
tool_6dof[0.42, −0.11, 0.63]
outcome● success
confidence 0.94
Prototype pipeline output on licensed stock footage · illustrative
Learning from experts across
- Automotive repair
- Industrial maintenance
- Food preparation
- Electronics assembly
- Surgical training
- Warehouse logistics
- Machining
- Construction trades
- Lab automation
- Home services
- Automotive repair
- Industrial maintenance
- Food preparation
- Electronics assembly
- Surgical training
- Warehouse logistics
- Machining
- Construction trades
- Lab automation
- Home services
01 · The insight
Every expert is a robot demonstration.
A mechanic repairing an engine. A chef preparing a meal. A surgeon performing a procedure. These people aren't just doing their jobs — they are generating demonstrations of how to interact with the physical world.
caliper · 0.94
02 / 05
Technician
Servicing industrial equipment
- 01pick up caliper
- 02open jaws
- 03measure bore
- 04log reading
knife · 0.98
03 / 05
Chef
Preparing a meal
- 01stabilize produce
- 02position blade
- 03slice ×12
- 04transfer to tray
iron_tip · 0.92
04 / 05
Assembler
Assembling a product
- 01pick component
- 02orient to pad
- 03insert
- 04solder & inspect
forceps · 0.95
05 / 05
Surgeon
Performing a procedure
- 01request instrument
- 02handoff
- 03precision grip
- 04return
Hover a card to see the demonstration it contains.
02 · What we extract
From raw pixels to robot-ready structure.
Our models transform raw egocentric video into structured representations designed specifically for training physical AI. Drag the divider.
hand_R · power_grasp
ratchet_3/8 · 0.96
socket_12mm · 0.93
engine_bay · ctx
RAW VIDEO
demo_000184.json● valid
"task": "loosen_bolt""domain": "automotive""subtasks": ["reach","grasp","seat","rotate"]"objects": ["ratchet_3/8","socket_12mm"]"hand": "right""grasp_type": "power_cylindrical""keypoints": [21, 3] // per frame"contact": { t: 15.2s, obj: ratchet }"tool_axis": [0.06, 0.99, 0.08]"rotation": -94.5° // ccw"trajectory": SE(3) × 212"success": true01
Tasks & subtasks
Hierarchical segmentation of long-horizon work into goals and steps.
02
Objects & tools
Open-vocabulary detection and tracking of every tool and part in play.
03
Hand & body motion
21-point hand pose and full-body kinematics, per frame.
04
Object interactions
Contact events, grasp types and force-bearing moments.
05
Temporal sequences
Ordered action graphs with timing, pauses and retries.
06
Actions & skills
Reusable skill primitives mapped to a shared taxonomy.
07
3D trajectories
Metric 6-DoF paths of hands and tools lifted from monocular video.
08
Success & failure
Labeled outcomes, including the mistakes experts recover from.
03 · The pipeline
From video to robot intelligence.
We turn unstructured video into the building blocks of robot learning.
- 01
Expert Video
Head-mounted and body-worn footage from people doing real work.
- 02
Perception
Hands, bodies, objects and depth, reconstructed from every frame.
- 03
Tasks & Skills
Long videos segmented into goals, subtasks and skill primitives.
- 04
Actions & Interactions
Contacts, grasps and state changes grounded in 3D.
- 05
Structured Demonstrations
Robot-ready trajectories in standard learning formats.
- 06
Robot Training
Policies and world models pre-trained on human experience.
04 · The data layer for physical AI
Robots shouldn't have to generate all their own training data.
Instead of asking robots to collect every demonstration themselves, we learn from the humans who already know how to do the work.
Today
Robot teleoperation
- Limited in scale
- Expensive to collect
- Tied to specific hardware & environments
What we're building
Human expert video
- Massive in scale
- Naturally diverse
- Captured across environments, tools & tasks
Order-of-magnitude comparison
Human expert video Robot teleoperation
Cost per hour of demonstration
lower is better
< $5
$100–$200
Distinct environments per 1k hrs
higher is better
1,000+
~10
Hours collectable per month
higher is better
100k+
~1k
Illustrative estimates from public reporting and our own early pilots. Not a guarantee of future performance.
05 · The vision
Millions of hours. Thousands of skills.
The long-term vision is a foundational dataset of human interaction with the physical world — not captions, but complete, grounded procedures.
Not just
“A person is holding a screwdriver.”
But
“A person identifies the correct fastener, reaches for a screwdriver, grasps it, positions it against the fastener, applies torque, verifies the result, and moves to the next step.”
action_graph · fasten_panelstep 0/7
gaze
fixate
track
inspect
hand_R
reach
grasp
position · torque
release
hand_L
stabilize part
tool
screwdriver_PH2
✓
t0t1t2t3t4t5t6
At scale, these trajectories become a library of reusable physical skills.
Pick up.
pick_up()align()insert()rotate()connect()tighten()inspect()assemble()repair()
06 · Beyond imitation
Not copying humans. Understanding how actions change the world.
By learning from enormous collections of expert trajectories, we can work toward models that predict the consequences of physical actions.
statesₜ
+
actionaₜ
→
future stateŝₜ₊₁
fθ( state, action ) → future statelearned world modelReason before acting
Robots imagine the outcome of an action before executing it.
Train in simulation
Learned environments let policies practice millions of times, safely.
Learn without demos
Acquire skills that were never demonstrated directly on a robot.
Video becomes the foundation for a learned environment.
07 · Early results
Early signal from our prototype pipeline.
We're early. These are preliminary numbers from an internal pilot — small, honest, and improving every week.
0+
Hours of egocentric video processed
0+
Skill primitives in our taxonomy
0%
Subtask segmentation F1
0 mm
Mean 3D hand-pose error
Internal benchmark · 150 held-out expert clips
score, 0–100 · higher is better · preliminary
ElectricAI pipeline (v0.3) Off-the-shelf VLM baseline
Subtask segmentation F1
91
64
Tool identification accuracy
94
78
Contact-event recall
87
52
Grasp-type classification
82
49
Internal evaluation on 150 held-out egocentric clips across 6 task domains. Baseline: a general-purpose vision-language model prompted zero-shot. Results are preliminary and not peer-reviewed.
08 · Team
A small team, obsessed with physical AI.
A technical founding team working at the intersection of computer vision, human motion and robot learning.
Robotics & computer vision. Previously building perception systems for autonomous vehicles.
09 · FAQ
Questions we hear often.
Robotics teams and physical-AI labs that need large, diverse demonstration data to pre-train manipulation policies and world models — without standing up a fleet of teleoperated robots.
Human expertise → Data → Intelligence → Robots
The physical world is the next training corpus.
We're partnering with robotics teams, data contributors and investors who want to build the foundation model for physical work.
alexyang@electric-ai.link