Get started
Step 3 / 5

Rebuilding the scene

Walls, doors, stairs

LiDAR0.0M pts

0 objects

found so far
Physical AI, Data and simulation

Upload footage or LiDAR. Force agents rebuild the space in 3D, add real physics and turn it into robot training data.

Get startedSee it work
From footage to a trained robot

Footage, Labels, 3D Simulations, physics, and training all in a few clicks. Force runs them as agents you can watch, check and stop.

Try it on your footage
1 Footage
Helmet-camera footage of two hands reaching for a mug on a kitchen counterREC 00:12:31CAM 3, Helmet, 1080p

Review footage

Downloading, clip 1 of 31

2 Labels
0:02 right, Reach for the mug
0:04 right, Grasp the mug by the handle
0:06 right, Lift the mug 20 cm
0:09 left, Steady the counter

Label hand actions

Starts after step 1

3 3D scene
5.0M points become 394 objects

Rebuild the scene

Waits for your yes

4 Training
Sim, Pick & place, not yetA robot arm lifting a mug

Train the robot

Starts after step 3

How it works1

3 steps: You answer a few questions in the chat and approve or reject anything that changes and let the Force agents do the rest.

1

Upload footage or LiDAR

Drop clips in, or point Force at S3, Google Cloud or your own server. Helmet cams, phones and LiDAR scans all work.

helmet_cam03_0412.mp4412 MB
kitchen_scan_02.e571.8 GB
helmet_cam01_0414.mp4266 MB
S3GCSSFTPLocal
2

Agents label, rebuild and add physics

Every hand action is labelled with a timestamp. The space is rebuilt in 3D at real scale, with mass, friction and joints.

both, Pick up the plate90%
Mass 0.42 kgFriction 0.6Hinge 1
3

Train and test the robot in sim

Describe the task in one sentence. Watch the robot try it next to the person who did it, then export to your simulator.

Success70.0%
Isaac SimMuJoCoUSD
See it work

Every action gets a color, a timestamp and a confidence score. The space behind it becomes a measured model a simulator can use.

Helmet-camera clip: a person's hands reach for a mug on a kitchen counterleft handright handMug, 0.32 kgCAM 3, Helmet, 0:00.0both, Hands at rest98%
0:00.0 / 0:18
IdleReachGraspMovePlaceReleaseHatched: sent to a person
Reconstructed 3D model of a small stone buildingDoor 1, 2.1 × 0.9 mBefore LiDAR, 7.85M pointsAfter Reconstructed mesh
Building 2, 394 objects, walls, doors, stairsDrag to compare
Use cases2

If a person can do it on camera, Force can turn it into data a robot learns from.

Scene 1, Kitchen

Helmet-cam footage of people cooking and cleaning becomes pick, place and pour demos.

actions labelled
51
objects
20
pick success
97.8%
LabelsReachGraspPourWipeOpen drawer
Playground1

Type the task. Force finds the moment in your footage, gives every object real physics and shows the robot doing it next to the person who did.

Command
>Pick the cup off the counter the way the
  • Found the grasp in clip 1 at 0:04Action labeller
  • Matched the mug, 0.32 kg, friction 0.6Physics estimator
  • Replayed the human demo in the kitchen sceneRobot gym
  • Trained across 500 variationsRobot gym
97.8%success in 500 tries
Ready to export, USD, MJCF
The person's right hand grasping the mug by its handle
Human, Clip 1, 0:04
A robot arm lifting the same mug in simulation
Robot, Sim, Try 0Sim, Pick & place, 97.8%
Grip force
not yet
Lift
not yet
Slip
not yet
Match to human
not yet

Works with

Bring your own footage or try the sample kitchen pack. Your first run is on us.

Get startedLog in