SIGNAL
Self-driving research has been quietly swapping a dangerous habit for a brittle one. For years, the only credible way to teach a car to drive was to let it drive, sometimes on real roads, sometimes in simulation, accumulating the long tail of mistakes until the policy converged. The newer instinct is to never let it drive at all: train on recorded logs, stay in the simulator, and trust offline data to do the work. That swap solves one failure mode and hands the field four more: distribution shift, heavy-tailed risk, out-of-distribution actions, and a state space full of redundancy the policy cannot filter on its own.
The DiDrive preprint attacks all four with the same move. It splits the problem in two. One half, a risk-gated encoder, decides which parts of the scene deserve attention and discards the rest. The other half, an in-sample calibration step, keeps the learned policy from drifting into actions the logged data never showed. In a 60-vehicle dense-traffic CARLA scenario, the authors report an 85% success rate and a 4295.68 average reward against IQL, CQL, and Diffusion-QL.
The honest read: this is one benchmark on one simulator, pre-peer-review, and offline learning's track record against deployment reality is mixed. But the mechanism is reusable. Anywhere a policy is asked to learn from someone else's driving logs, the fix is the same: separate perception from action, and refuse to act on what the data has not yet seen.