FlowMo-WM: New world model handles momentum and hidden drift for robots
Robots can now navigate water currents and inertia without flow sensors.
FlowMo-WM addresses a key gap in robot learning: most world models assume motion is dominated by immediate control, ignoring inertia and hidden ambient drift like water currents or wind. This paper introduces an end-to-end trainable visual world model that infers object-centric motion state and a predictive long-history context from image-action pairs—without direct flow supervision. It splits history into a short-term latent state for motion and a longer-term context for exogenous factors, then uses a zero-context residual transition to separate action-driven base dynamics from drift effects during latent rollout.
In simulated aquatic surface-vehicle environments with diverse hidden flows and randomized dynamics, FlowMo-WM significantly boosts long-horizon prediction accuracy compared to standard action-conditioned latent world models. Context ablations (zeroing or shuffling the inferred context) prove the ambient context is critical for stable prediction under drift. This work, from Yitao Jiang, Luyang Zhao, Muhao Chen, and Devin Balkcom, is a step toward more realistic robot controllers for real-world autonomous boats, drones, and other vehicles affected by unseen environmental forces.
- FlowMo-WM factors image-action history into short-term motion state and long-term context for hidden drift.
- Zero-context residual transition separates action-driven dynamics from drift effects during rollout.
- In simulated aquatic environments, it improves long-horizon prediction accuracy over existing latent world models.
Why It Matters
Enables robots to predict motion accurately in uncertain environments like rivers or winds without expensive flow sensors.