doPlan: A Variable-Horizon Dataset for Multi-Stage Language-Conditioned Planning in Autonomous Driving
The dataset is built to test whether driving planners can hold passenger intent across delayed, conditional, and multi-step road events.
doPlan adds 5,154 human-written passenger instructions on top of nuPlan, spanning 50.9 hours of unique real-world driving. Its annotation windows run from 30.0 to 508.8 seconds, covering intent that may be immediate, deferred, event-conditioned, persistent, or multi-stage. In the authors’ evaluation, four language-conditioned driving models did not consistently turn language sensitivity into behavior aligned with the requested direction. Among matched future maneuvers, only 9.8% occurred within a common 5-second prediction horizon. Source: ArXiv · AI/CL/LG's note.
doPlan adds 5,154 human-written passenger instructions on top of nuPlan, spanning 50.9 hours of unique real-world driving. Its annotation windows run from 30.0 to 508.8 seconds, covering intent that may be immediate, deferred, event-conditioned, persistent, or multi-stage. In the authors’ evaluation, four language-conditioned driving models did not consistently turn language sensitivity into behavior aligned with the requested direction. Among matched future maneuvers, only 9.8% occurred within a common 5-second prediction horizon. Source: ArXiv · AI/CL/LG's note.
score 4