Velox World · Research team

World models for physical AI.

We build efficient world models: systems that learn how the physical world looks, moves and responds to action, and run fast enough to plan and act inside it.

Built toward

Explore the research Our vision

Generated, not rendered Samples from our models

01 · Vision

Physical AI needs a world it can rehearse in.

Robots, vehicles and simulators act in a world where trying everything is slow, costly or unsafe. A world model lets an agent imagine what happens next before it commits. That only works if the imagination is consistent over long horizons, trained toward what matters, and fast enough to sit inside the control loop.

01

Consistent

Stay coherent as the horizon grows: remember what has been seen, and recognise it when the camera or the agent comes back.

Velox →
02

Aligned

Be trained toward the signals that matter, with reinforcement learning cheap enough that we can actually iterate on it.

RPO →
03

Real-time

Meet a latency budget. Memory, attention and training all have to be efficient before a world model can be used inside a loop.

Both

ImagineCompareActinside the loop

02 · Research

Making world models cheap enough to run live.

Real-time interaction is both a systems problem and a learning problem. Our first two projects work on each end: how a world model remembers at inference time, and how a visual generator is aligned at training time.

DenseVelox

Velox · Consistent

Spatial memory makes long‑context world models efficient

A training-free attention router. Camera frustum slabs from streaming geometry tell each new region of the video which parts of the past actually matter, so later chunks stop paying for the entire history.

4.36×attention · chunk 48

2.43×denoising · chunk 48

0training steps

Open the Velox project

Base+ RPO

RPO · Aligned

Routing Policy Optimization for online diffusion RL

Attention routing becomes a discrete policy. Two matched rollouts differ only in their sampled routes, so one reward difference is enough to learn, and inference goes back to dense attention with the original sampler.

24.1GPU-hours · Wan2.2

3.1–4.2×less than GRPO

+3HPSv3 · dense eval

Open the RPO project

03 · Direction