Build AI World Models for Real-Time Decision Making
The state vector is the most underrated design decision. Start with sensor readings, then remove redundancy (e.g., wheel currents are often correlated), and finally add back the hidden variables that cause hysteretic effects (battery internal resistance, gearbox wear). In practice I iterate over the vector while plotting real trajectories until I can predict 3–5 steps ahead without drift.
Don’t skip the KL term. Without it the network tends to ignore the latent state and falls back to a Markovian predictor, which drifts after 2–3 steps.
Horizon RMSE is measured against held-out episodes. INT8 adds < 10 % error while cutting memory and latency by > 2×. On ARM Cortex-A72 (Raspberry Pi 4) INT8 inference drops to 12 ms, meeting real-time requirements.
Deploy early and often, even with stub predictors (constant trajectory, random perturbations). Instrument full state logging from day one; you’ll need those logs to debug why the model behaved badly in production.