MomADv2, a preprint autonomous-driving planner, reports its clearest result at a six-second planning horizon: an average L2 displacement error, or path-distance error, of 1.21 metres and an average collision rate of 0.76% on open-loop nuScenes validation. The authors say those figures represent relative improvements of 11.7% and 11.6%, respectively, over the strongest prior method. The abstract also reports a 15.6% relative reduction in average collision rate compared with MomAD under six-second planning.
That distinction matters when reading the headline percentage. A 15.6% relative reduction is not the same as a 15.6-percentage-point drop. The absolute collision-rate figure reported for the six-second nuScenes test is 0.76%, while the accompanying average L2 error is 1.21 metres. In the paper's comparison, MomADv2 is the lowest-error and lowest-collision-rate entry for that open-loop validation result.
The model filters memory before refining a path
MomADv2's design is built around information from earlier planning steps. It combines a Selective State-Space Planning Memory Query Module, which filters historical planning queries, with a Flow-Matching Trajectory Residual Refiner that corrects planned trajectories toward expert trajectories. The first component uses predefined driving commands when selecting memory, a dependency the authors identify as a limitation. The refiner then provides the reported trajectory-correction step.
Training followed MomAD pretraining and then module adaptation. The second stage ran for 10 epochs at a learning rate of 3 x 10^-6 on eight NVIDIA RTX 4090 GPUs. Most pretrained perception, motion and planning components were frozen during that stage. The comparison therefore centers on the added memory and refinement modules within that stated configuration.
The evaluation was spread across several benchmarks rather than a single pooled sample. The reported nuScenes set contained 1,000 driving scenes and used multi-view images. Bench2Drive used a 1,000-clip base split and a separate 220-route evaluation set. The paper also reports NAVSIM settings, with nuScenes evaluated open loop and Bench2Drive evaluated closed loop, so each score should be read within its stated dataset and protocol.
Results varied by benchmark
Bench2Drive showed higher results in the paper's direct comparison with MomAD. MomADv2 recorded a Driving Score, or DS, of 52.32 against MomAD's 47.91. Its Success Rate was 24.24%, compared with 18.11% for MomAD. Elsewhere in the Bench2Drive results, MomADv2 is reported at 78.82 DS and 31.23 Comfortness. The paper presents these as separate reported results, so the values should not be treated as one combined score.
NAVSIM results covered multiple versions and splits. On NAVSIMv1 navtest, MomADv2 posted NC 99.0, DAC 97.1, TTC 95.5, EP 83.2 and PDMS 89.9. On the NAVSIMv2 navhard split, its overall listed EPDMS was 39.5. On NAVSIMv2 navtest, EPDMS was 87.9, and the model ranked first on DDC, EP and EC. The paper therefore reports strong results across several NAVSIM measures, but the numbers remain tied to their individual versions and splits.
Ablations point to a moderate history
The ablation results offer a closer look at how the reported system was assembled. In the module comparison, PDMS was 84.0 for the base configuration, 86.9 with SSM-Q and 89.9 with SSM-Q plus FM-Ref. A separate memory-strategy comparison found that the full selective-memory strategy had the lowest reported collision rate on nuScenes validation, at 0.76%. A history length of K = 4 achieved PDMS 89.9 on NAVSIMv1 navtest, while longer histories gradually degraded performance.
A separate trajectory-consistency analysis reported a 6.6% lower average TPC error for MomADv2 than for MomAD. Its reported TPC error was 1.04 metres at four seconds, 1.32 metres at five seconds and 1.61 metres at six seconds. The paper therefore reports consistency at three planning horizons, not only at the six-second endpoint.
The evidence has a clear boundary
The main limitation identified by the authors concerns the information used to choose memory. MomADv2 relies on predefined driving commands for that selection and does not explicitly model uncertainty in how future scenes may evolve. That caveat means the reported selective-memory result is not a command-free or uncertainty-aware memory test. The manuscript is labeled arXiv:2608.23405v1 [cs.CV] and dated 24 Aug 2026, making this a preprint report of benchmark results.
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