MomADv2, a model for autonomous-driving planning, reported lower long-horizon errors than comparison systems in benchmark tests. On the open-loop nuScenes evaluation, its average L2 displacement error was 1.21 metres and its collision rate was 0.76% when planning six seconds ahead. Against the strongest prior method reported in that comparison, the paper gives relative improvements of 11.7% for L2 error and 11.6% for collision rate.
Why the past matters
The abstract gives a second headline comparison with MomAD. Under six-second planning across the reported benchmark settings, it says MomADv2 had a 15.6% lower average collision rate than MomAD. The paper presents that figure as a relative reduction, so it should be read as a comparison between rates rather than an absolute collision percentage.
The model is aimed at a specific weakness in temporal planning: historical information can become unreliable when the driving command changes. MomADv2 combines a Selective State-Space Planning Memory Query Module, which filters history according to temporal continuity and command consistency, with a Flow-Matching Trajectory Residual Refiner, which learns residual corrections toward the expert trajectory. In practical terms, the design keeps history that remains consistent with the current command and suppresses history that does not, then applies a learned correction to the trajectory.
Three ways to test the model
The tests covered three named evaluation settings: open-loop nuScenes, closed-loop Bench2Drive and NAVSIM. The described data included 1,000 nuScenes driving scenes, a 1,000-clip Bench2Drive base split and 220 Bench2Drive evaluation routes. The reported measures were L2 displacement error and collision rate for nuScenes, Driving Score and Success Rate for Bench2Drive, and a closed-loop planning assessment for NAVSIM.
The gains were not confined to one benchmark
Bench2Drive produced a separate set of reported scores. MomADv2 had an average L2 of 0.77, a Driving Score of 78.82 and a Success Rate of 46.50%, alongside Efficiency of 83.41 and Comfortness of 31.23. The text also compares MomAD figures of 47.91 and 52.32 for Driving Score and 18.11% and 24.24% for Success Rate.
NAVSIM results extended the comparison. On NAVSIMv1 navtest, MomADv2 reported NC of 98.7, DAC of 96.8, TTC of 95.4, Comfort of 100 and EP of 84.5, with a PDMS of 89.9. The paper says that PDMS tied DriveSuprim for the best score and exceeded TransFuser by 5.9 points. On NAVSIMv2, it reported EPDMS of 39.5 on navhard and 87.9 on navtest, and ranked first on DDC, EP and EC on navtest.
Keeping plans consistent
Another test focused on whether plans stayed consistent as the horizon lengthened. MomADv2's trajectory prediction consistency, or TPC, errors were 1.04 metres at four seconds, 1.32 metres at five seconds and 1.61 metres at six seconds. Compared with MomAD, the paper reports relative improvements of 12.6% and 9.0% at the four- and five-second horizons, with a 6.6% average reduction in TPC.
What the model's own tests showed
The authors also varied the model's components and memory settings. In a NAVSIMv1 ablation, PDMS was 84.0 without the added modules, 86.9 with the selective memory module alone, and 89.9 with both the memory module and trajectory refiner. In the memory tests, full selective memory had the lowest reported collision rate at 0.76%, while a history length of K = 4 gave the best PDMS, 89.9. Longer histories degraded performance in that comparison.
The reported scope
MomADv2 is presented in an arXiv preprint, version 2, dated 30 September 2026. The study reports its figures on the named nuScenes, Bench2Drive and NAVSIM evaluation protocols, so the evidence presented here is a set of benchmark model comparisons across those settings.
Paper data and sources
Original title: MomADv2: Reliable Temporal Memory for End-to-End Autonomous Driving
Authors: Ziying Song, Shengkai Zhang, Lin Liu et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-24
DOI: Not available
Original paper · Full text