Every driving scene is logged once, but driving is not one-to-one. The same scene makes many behaviors feasible, and a language instruction picks among them. A standard log, however, records only the action the driver actually took. That leaves no direct supervision for the other reasonable actions, which is a problem for training models that must follow varied instructions.
The arXiv paper "Cartender: Verifiable Language-Conditioned Driving by Learning the Drives Not Taken" proposes to address this by learning from the drives not taken. Rather than treating the logged trajectory as the only correct answer, the method appears to model the space of feasible behaviors and use language to select within it. The abstract is truncated, so details of the architecture and experiments are not available in the source.
The core idea is that verifiability comes from understanding what else could have happened. If a model can distinguish between feasible and infeasible actions, and then map language to the right feasible action, it can be checked against the scene rather than only against a single recorded path. This shifts the learning signal from one observed outcome to the structure of possible outcomes. The source does not provide results or comparisons, so the effectiveness remains to be seen.