World models for flight
Learning to predict what a camera will see next, in latent space, and planning against that prediction. Image goals replace coordinates, so there is nothing for jamming or spoofing to take away.
One claim, reached in five phases: a latent world model plans a real quadrotor to camera-image goals, onboard, in real time, with no GPS. Each phase ships a measurable result before the next begins.
Model proposes. Deterministic code disposes. Three learned layers sit above a flight controller that no model ever touches.
A planner reads what the aircraft can do and breaks a goal into skills. A world model executes each skill by planning in latent space from a camera-image goal: roll out candidate actions, execute the first, look again, re-plan. A typed-decision gate answers bounded questions about the state with a probability and a confidence. Below all of it, attitude and position control stay classical.
Predicts what the camera will see next in representation space, never in pixels. Goals are images, so the aircraft needs no external positioning to know where it is going.
Bounded questions, typed answers, calibrated confidence. Low confidence means slow down, look again from another angle, or hand the decision to a person.
The backbones we use are months old and will be replaced. Every interface is model-agnostic so that nothing in the stack is married to one of them.
The flight controller is untouched. Safety limits, geofences and preconditions are code that every proposed action must pass before it runs.
The four questions the lab is working on, and why each one matters in the air.
Learning to predict what a camera will see next, in latent space, and planning against that prediction. Image goals replace coordinates, so there is nothing for jamming or spoofing to take away.
Sampling-based planning takes seconds per action. We distil planners into policies and shrink the sampling budget, and report the latency-versus-success trade-off with a bound, not a demo.
A model that says 0.8 should be right eight times in ten. We measure calibration against ground truth, test for bias under perturbation, and route low-confidence cases to a person.
Everything runs on Jetson-class hardware on the airframe. No tether, no ground-station GPU, no network in the flight loop, because the network is the first thing to go.
Razbotics is a deep-tech research lab in India working on visual world models for small uncrewed aircraft. We build the autonomy layer on commodity airframes and keep classical flight control untouched. The target is GPS-denied navigation and decision support for defence, reached through published research and India's defence innovation programmes.
The generic capability stays open: code, checkpoints, flight data and the evaluation protocol will be released so others can reproduce the result. Defence-specific layers stay closed. We decide which side of that line every artifact falls on before it leaves the lab.
For defence organisations, research partners, universities, and investors. Tell us who you are and what you are working on, and we will come back within two working days.