Most self-driving planners either recompute a fresh route every fraction of a second or stick stubbornly with an old one. A new system tries to do neither by default, and only step in when the road actually gets riskier.
The framework, called RiskWorld, comes from a paper posted to arXiv (arXiv:2609.18442) on September 17, 2026. It fuses spatial risk fields and short-term actor context with bird's-eye-view camera features, then uses flow-guided evolution to forecast how occupied space around the car will change, correcting that forecast with learned residuals. Crucially, it generates one forecast per planning step and reuses it across every candidate route, rather than rerunning prediction for each option. A candidate only replaces the car's current plan if it clears a nonnegative collision-score threshold against a do-nothing baseline and meets separate limits on predicted risk and trajectory error. Tested open-loop on the nuScenes driving dataset, RiskWorld posted the lowest collision rate at a 3-second horizon and the second-best average trajectory error among the compared baselines, while running at 11.5 FPS on a single RTX 4090 with about 91 million parameters.
The interesting part isn't the forecasting model itself - bird's-eye-view occupancy prediction is well trodden ground. It's the gatekeeping: instead of treating every new prediction as a reason to replan, RiskWorld treats the current trajectory as innocent until proven risky, and reuses one forecast across many candidates to keep that judgment cheap. That's a real answer to a real production problem, since constant replanning burns compute a car doesn't have to spare.
Worth remembering: this is an open-loop benchmark result, not a road test, and nuScenes has flattered plenty of planners that struggled once traffic got messy. Selective replanning is a sound instinct. Whether it holds up outside a curated dataset is a separate question the paper doesn't answer.