World models that take robots into homes, factories and streets
KeepOne builds video-native world models for robots. The robot first imagines what will happen next, then decides how to act. Our DeltaToken representation cuts video tokens by two orders of magnitude, which makes pre-training cheap, on-device inference real-time, and minutes of memory practical.
Real videoGenerated by DeltaWorldPrompt: Unfold the towel and fold it neatlyego · 5s
192×fewer video tokensvs Wan VAE · 480×832 · 24fps · 5s
976tokens for a 5-second clipa VAE needs 187,200
5minof long-horizon memorynew tasks via in-context learning
4mmthick visuo-tactile sensorothers: 15–20mm
Next demos
In the coming months we will ship end-to-end demos in three real-world settings
All three run on the same world model. Next to each one is the part we already have working today.
HomeIn development
A Unitree R1 humanoid doing chores in a real home
Give one spoken instruction and the robot completes long, multi-step chores in a home it has never seen. The world model understands the scene and rehearses the motion before the body executes it.
Robot
Unitree R1 humanoid
Tasks
Folding towels, tidying up, wiping furniture, opening the washing machine
Goal
Natural-language instruction to finished chore, end to end, without per-task teleoperation data
Video + hand pose
TodayDeltaWorld generates future video and both hands' pose trajectories from one instruction: “Fully open the washing machine door using both hands.”
FactoryIn development
Tabletop manipulation at 99% success that generalizes to new parts and new stations
Factories buy reliability. DeltaSim turns a single photo into a simulated workcell so we can evaluate policies systematically and find out why they fail. Tactile feedback closes the loop on the contact that vision cannot see.
Robot
Tabletop arms with our visuo-tactile sensors
Tasks
Picking, sorting and placing
Goal
99% success; new objects, layouts and lighting without collecting new data
99%target task success rate
RealSim
TodayDeltaSim rebuilds an interactive simulation from one photo of a real scene, and the same policy runs side by side in both.
OutdoorIn development
Fully autonomous delivery from point A to point B
The goal is a robotaxi for food and parcel delivery. A delivery robot leaves the pickup point, crosses real streets and reaches the customer with no human takeover.
Robot
Outdoor delivery robot
Tasks
Last-mile food and parcel delivery
Goal
A to B, fully autonomous, from video directly to control
A → Bzero human takeovers
RealGenerated
Delivery data capture
TodayDeltaWorld already generates driving video, and DeltaSource is collecting first-person data from food and parcel couriers at scale.
DeltaToken
How many tokens does a 5-second video need?
Token count is the real cost curve of a world model. DeltaToken borrows from video codecs: it encodes only what changes between frames, as a 1D causal stream of tokens, and a diffusion decoder rebuilds the pixels.
976 tokens1 square = 100 tokens
Wan VAE187,200
Wan VAE (2×2 patch)46,800
DeltaToken976
480×832 · 24fps · 5s. Wan VAE = 480/8 × 832/8 × 24/4 × 5. Both use 16 channels.
100×+
cheaper training and inference, since transformer compute grows with the square of token count.
100×
longer video in the same token budget, which makes long memory and long-horizon tasks possible.
$3–4k
to train a 1.6B DiT prototype on 3.2M clips, in 32 B200 GPU-days.
System · S3 → S0
Four layers, from planning every few seconds to control a thousand times a second
S3Planner0.1–0.3 Hz
A harness agent built on tools like Claude Code, with a memory module, plans the task at the top level.
S2Slow system · World model1–2 Hz
Reasons about and generates the next video and actions, with video prompting and video thinking. A Mixture-of-Transformers model unifies understanding and generation in the DeltaToken space.
S1Fast system · Inverse model20–30 Hz
Translates S2's latent plan into concrete actions, using body state and sensors including touch. One is trained per robot body.
S0Controller500–1000 Hz
Turns S1's actions into motor currents. Our own motion control removes singularities and makes motion smoother.
Full stack
Five modules, from data to simulation, in one loop
DeltaSource
A consumer app that crowdsources first-person video from couriers, homes and factories.
DeltaData
A streaming pipeline that filters, segments by atomic action, captions, rebalances and pre-computes.
DeltaToken
A streaming video tokenizer that encodes only frame-to-frame change, 100× more efficient than a VAE.
DeltaWorld
A world model that unifies video understanding and generation, and predicts future video and actions.
DeltaSim
Builds embodied simulations from a single image or video, for evaluation, failure analysis and RL.
Failures found in DeltaSim tell DeltaSource what data to collect next.
Results today
What already works
Every clip below comes from our own models and systems.
Household · top real / bottom generatedWipe the fitness equipment seat with a brown towelHousehold · top real / bottom generatedPut the black box back into the transparent plastic box and close the lidWarehouse · top real / bottom generatedSort and place the small cardboard box on the shelfHousehold · top real / bottom generatedUnfold the towel and fold it neatlyDriving · top real / bottom generatedDaytime road sceneDriving · top real / bottom generatedNight-time city street
Top video / bottom posePick up the towel with your right hand.Top video / bottom poseCover the pot with the lid using your left hand.Top video / bottom poseFully open the washing machine door using both hands in coordination.
Pose trajectories are generated together with the video and serve as an action prior for the S1 inverse model.
Left real / right simPick up one of the water bottles with your left hand and throw it into the trash can.Left real / right simPick up the snack on the second shelf from the left with your left hand.Left real / right simPick up the TV remote with your right hand and place it in the bedside table drawer.Failure analysis · new camera angleRe-rendering from another angle shows the cause: the gripper did not squeeze hard enough. The fix is closed-loop control or a tactile sensor.
DeltaSource · food & parcel deliveryFirst-person video from real delivery runsDeltaSource · householdStereo first-person capture of chores
Data pyramid: internet and monocular ego video for pre-training, stereo ego with wrist views for mid-training, tactile and teleoperation data for post-training.
Our visuo-tactile sensor: 15 × 15 × 4 mm
It measures both normal and shear force with high precision, and it is small enough to cover a whole palm. We build the sensor, so we own the tactile data.
KeepOne visuo-tactile sensor4 mm
Thinnest alternatives15–20 mm
Drawn to scale by thickness, full scale 20 mm.
Live tactile readoutTactile signal while pressing the sensor
Roadmap
Roadmap
Done
Core technology proven
DeltaToken: 192× fewer tokens at reconstruction quality comparable to a VAE
DeltaWorld: future video and pose trajectories from an instruction
DeltaSim: an embodied simulation from a single image
4 mm visuo-tactile sensor prototype
December 2026
Foundation model and cross-embodiment
Next version of DeltaToken
World model pre-training complete
Inverse models for multiple robot bodies
Zero-shot transfer across bodies
Thinking and reasoning in the model
Coming months
End-to-end demos in three settings
Home: Unitree R1 doing chores
Factory: tabletop tasks at 99% success
Outdoor: fully autonomous A-to-B delivery
Bring world models into the real world with us
We are looking for deployment partners, robot makers and investors. We are also hiring researchers and engineers in world models, embodied RL, robot systems and tactile sensing.