Latency is learnable
A real-time model should not aim at where the world was when computation began. Horizon-Min learns to predict where the target will be when an action can actually arrive.
Latency changes the correct answer. If a perception system needs 300 milliseconds to decide, the target may have moved before the action begins. Horizon-Min treats that delay as part of the prediction problem.
- Four sequential frames estimate motion instead of treating an image as timeless.
- The model learns a prediction horizon that adapts to its own measured processing delay.
- Training covers latency values from 100 to 600 milliseconds.
- In the controlled task, the adaptive system reaches a 60.9% hit rate.
- Random and fixed-center baselines remain near 5% and 8%.
Delay is part of the world
Real-time perception is usually described as a race to reduce inference latency. That is necessary, but it is incomplete. Some delay always remains: sensing, transfer, inference, planning, actuation. A system that ignores that delay may produce a precise answer for a moment that has already passed.
In interception tasks, the difference is visible. A model sees a moving object, estimates its current position, and sends an action. By arrival time, the object is elsewhere. The useful target is not the current coordinate. It is the future coordinate at the end of the system's own action horizon.
Four frames, then a lead
Horizon-Min observes four sequential frames, estimates velocity, and predicts a lead point. The lead is not fixed. It depends on the measured latency of the current processing path.
The training procedure varies that latency between 100 and 600 milliseconds. This forces the model to connect delay with the amount of lead required. A slower inference pass should result in a farther prediction horizon. A faster pass should not overcompensate.
The current system is vision-language-model first. The broader direction is vision-language-action, where the same temporal reasoning can sit closer to the action policy.
The controlled result
| System | Hit rate | What it represents |
|---|---|---|
| Adaptive horizon | 60.9% | Lead changes with measured latency |
| Fixed center | About 8% | No motion-aware lead |
| Random | About 5% | Chance baseline |
The gap shows that timing information is not merely a systems detail in this task. It is predictive input. The agent can learn how far ahead to act when its own delay varies.
What this does not prove
The reported numbers come from a controlled interception environment. They do not establish the same hit rate in robotics, open-world interfaces, or safety-critical control. Those settings add uncertain sensing, nonlinear motion, action dynamics, and consequences that the present study does not model.
Horizon-Min is best read as evidence for a design principle: systems that act in time should represent their own delay inside the task. The next work is to test whether that principle survives richer motion, different model families, and action policies that learn more than a lead coordinate.