MomWorld: Momentum-Aware Latent World Model for Long-Horizon Autonomous Driving

Ziying Song; Shengkai Zhang; Lei Yang; Haozhuang Chi; Yuchen Liu; Jiangtao Su; Lin Liu; Ziyang Liu

cs.ROSep 27, 2026

MomWorld: Momentum-Aware Latent World Model for Long-Horizon Autonomous Driving

Authors: Ziying Song, Shengkai Zhang, Lei Yang, Haozhuang Chi, Yuchen Liu, Jiangtao Su, Lin Liu, Ziyang Liu, +1 more

Organizations: Nanyang Technological University · Beijing Jiaotong University · North University of China · Dalian University of Technology · Tsinghua University

Abstract

Long-horizon planning enables autonomous vehicles to anticipate scene evolution and potential risks, supporting safe and stable decisions in complex interactions. However, existing methods struggle to propagate motion trends from observed history into the future. Long rollouts based on a single latent state may further attenuate useful dynamics, retain stale motion patterns, and disrupt reliable near-term plans. We introduce MomWorld, a momentum-aware latent world model for long-horizon planning. MomWorld extracts scene motion trends from historical-to-current observations and propagates latent momentum into future horizons, jointly predicting future configuration and momentum states. A learnable momentum persistence mechanism preserves stable trends, scene-conditioned momentum updates adapt future dynamics, and a scene-adaptive reset gate suppresses stale momentum under abrupt changes. We further propose MoFlow, a momentum-conditioned flow-matching module that refines a base trajectory to align with the predicted future scene evolution in only a few integration steps, with a horizon-aware residual fusion that preserves near-term planning stability while permitting stronger long-range corrections. Extensive experiments on NAVSIM, nuScenes and Bench2Drive demonstrate that MomWorld improves long-horizon planning consistency and reduces the average collision rate by 12.2% relative to MomAD over a 6-second planning horizon.

Figures & tables

Figure 1: Motivation and comparison of MomWorld. (a) Reliable future scene evolution remains a key challenge for long-horizon planning. (b) MomAD ( Song et al., 2025 ) derives ego-centric momentum from historical and current evidence to stabilize ego dynamics. (c) MomWorld elevates momentum to a world-level latent state, initializes it from past-to-present evidence, and jointly rolls out future configuration and momentum to model world dynamics. (d) On six-second nuScenes planning, MomWorld reduces L2 from 2.45 to 2.31 m and collision rate from 2.13% to 1.97%. The corresponding averages improve from 1.42 to 1.17 m and from 0.90% to 0.79%.

Figure 2: Overview of MomWorld . Historical and current multi-view images are encoded into Scene Queries, from which MoLWM initializes latent scene state and momentum and propagates them through Latent World Rollout (LWR) to construct Future World Memory for candidate scoring and Plan selection. Guided by history-conditioned momentum and predicted future evolution, MoFlow applies bounded residual refinement to produce the temporally consistent Refined trajectory $\widetilde{\tau}$ .

Figure 3

$\operatorname{Method}$	$\operatorname{Venue}$	$\operatorname{L2\ (m)}\downarrow$							$\operatorname{Col.\ Rate\ (\%)}\downarrow$							$\operatorname{FPS}$
$\operatorname{Method}$	$\operatorname{Venue}$	1s	2s	3s	4s	5s	6s	$\operatorname{Avg.}$	1s	2s	3s	4s	5s	6s	$\operatorname{Avg.}$	$\operatorname{FPS}$
$\operatorname{UniAD}$ ( Hu et al., 2023 )	CVPR’23	0.47	0.91	1.35	1.91	2.47	3.07	1.70	0.25	0.36	0.61	0.99	1.64	2.51	1.06	1.8
$\operatorname{SparseDrive}$ ( Sun et al., 2025b )	ICRA’25	0.43	0.87	1.23	1.75	2.32	2.95	1.59	0.19	0.31	0.56	0.87	1.54	2.33	0.97	9.0
$\operatorname{MomAD}$ ( Song et al., 2025 )	CVPR’25	0.41	0.85	1.13	1.67	1.98	2.45	1.42	0.17	0.30	0.54	0.83	1.43	2.13	0.90	7.8
$\operatorname{LAW}$ ( Li et al., 2025b )	ICLR’25	0.40	0.87	1.16	1.71	2.03	2.61	1.46	0.19	0.33	0.57	0.86	1.51	2.31	0.96	19.5
$\operatorname{Epona}$ ( Zhang et al., 2025b )	ICCV’25	0.39	0.91	1.17	1.73	2.02	2.75	1.50	0.14	0.18	0.45	0.74	1.48	2.23	0.87	–

Table 1: Six-second planning results on the nuScenes validation set. Reported FPS uses an A100 for UniAD, an RTX 3090 for LAW, and an RTX 4090 for SparseDrive, MomAD, and GuideFlow.

$\operatorname{Method}$	$\operatorname{Venue}$	$\operatorname{TPC\ (m)}\downarrow$
$\operatorname{Method}$	$\operatorname{Venue}$	4s	5s	6s	$\operatorname{Avg.}$
$\operatorname{UniAD}$ ( Hu et al., 2023 )	CVPR’23	1.49	1.81	2.41	1.90
$\operatorname{VAD}$ ( Jiang et al., 2023 )	ICCV’23	1.55	1.73	2.17	1.82
$\operatorname{SparseDrive}$ ( Sun et al., 2025b )	ICRA’25	1.33	1.66	1.99	1.66
$\operatorname{MomAD}$ ( Song et al., 2025 )	CVPR’25	1.19	1.45	1.61	1.42
$\operatorname{MomWorld\ (Ours)}$	–	0.93	1.19	1.46	1.19

Table 2: Trajectory Prediction Consistency at 4–6 s on the nuScenes validation set.

Table 7

$\operatorname{MoLWM}$			$\operatorname{MoFlow}$		$\operatorname{nuScenes\ 6s}$					$\operatorname{NAVSIM\ v2\ navhard}$
FWM	LWR	SMD	MGRF	HARF	$L_{2}@3\,(\mathrm{m})\downarrow$	$L_{2}@6\,(\mathrm{m})\downarrow$	$\operatorname{TPC@6}\,(\mathrm{m})\downarrow$	$\operatorname{Col.@3}\,(\%)\downarrow$	$\operatorname{Col.@6}\,(\%)\downarrow$	$\operatorname{LK}\uparrow$	$\operatorname{EP}\uparrow$	$\operatorname{EC}\uparrow$	$\operatorname{EPDMS}\uparrow$
					1.13	2.45	1.61	0.54	2.13	54.6	69.5	49.7	41.7
✓					1.07	2.42	1.57	0.51	2.09	54.8	72.3	50.8	42.0
✓	✓				1.01	2.39	1.54	0.48	2.06	55.0	75.6	52.1	42.2
✓	✓	✓			0.96	2.36	1.51	0.46	2.03	55.1	78.4	53.0	42.4
✓	✓	✓	✓		0.91	2.33	1.48	0.44	2.00	55.2	80.2	54.0	42.6

Table 6: Roles of Different Method Components in MomWorld. Cumulative component study on the nuScenes validation set and the NAVSIM v2 navhard split. FWM, LWR, SMD, MGRF, and HARF denote Future World Memory , Latent World Rollout , Scene-Adaptive Momentum Dynamics , Momentum-Guided Residual Flow , and Horizon-Aware Residual Fusion , respectively. LK, EP, and EC are the Stage-2 metrics reported by the navhard protocol.

$\operatorname{Variant}$	$\operatorname{nuScenes\ 6s}$					$\operatorname{NAVSIM\ v1\ navtest}$	$\operatorname{NAVSIM\ v2\ navhard}$
$\operatorname{Variant}$	$L_{2}@3\,(\mathrm{m})\downarrow$	$L_{2}@6\,(\mathrm{m})\downarrow$	$\operatorname{TPC@6}\,(\mathrm{m})\downarrow$	$\operatorname{Col.@3}\,(\%)\downarrow$	$\operatorname{Col.@6}\,(\%)\downarrow$	$\operatorname{PDMS}\uparrow$	$\operatorname{EPDMS}\uparrow$
No historical variation	0.95	2.58	1.72	0.55	2.40	88.6	39.7
No horizon embedding in SMD	0.90	2.43	1.57	0.47	2.12	89.5	41.3
Fixed retention ( $\rho_{k}\equiv 0.9$ )	0.92	2.49	1.63	0.50	2.21	89.0	40.8
No momentum proposal ( $u_{k}\equiv 0$ )	0.99	2.69	1.82	0.60	2.58	87.8	38.6
No reset gate ( $g_{k}\equiv 0$ )	0.94	2.55	1.69	0.65	2.83	88.2	37.9

Table 7: Ablation study of different designs in MoLWM across nuScenes , NAVSIM v1 navtest , and NAVSIM v2 navhard .

$\operatorname{MoFlow\ Design}$	$\operatorname{Steps}$	$\operatorname{nuScenes\ 6s}$					$\operatorname{NAVSIM\ v1\ navtest}$	$\operatorname{NAVSIM\ v2\ navhard}$	$\operatorname{Latency\ (ms)}\downarrow$
$\operatorname{MoFlow\ Design}$	$\operatorname{Steps}$	$L_{2}@3\,(\mathrm{m})\downarrow$	$L_{2}@6\,(\mathrm{m})\downarrow$	$\operatorname{TPC@6}\,(\mathrm{m})\downarrow$	$\operatorname{Col.@3}\,(\%)\downarrow$	$\operatorname{Col.@6}\,(\%)\downarrow$	$\operatorname{PDMS}\uparrow$	$\operatorname{EPDMS}\uparrow$	$\operatorname{Flow}$	$\operatorname{Total}$
No MoFlow (base plan $\tau^{0}$ )	0	0.96	2.36	1.51	0.46	2.03	89.7	42.4	0.0	130.5
Unconditioned MGRF	4	1.01	2.53	1.67	0.57	2.38	88.3	40.0	8.4	138.9
$M_{t}$ only	4	0.91	2.40	1.55	0.47	2.09	89.4	41.8	8.4	138.9
$p_{0}$ only	4	0.93	2.43	1.58	0.49	2.14	89.2	41.5	8.4	138.9
$p_{0},M_{t}$	1	0.91	2.38	1.53	0.46	2.07	89.8	42.3	2.2	132.7

Table 8: Ablation study of different designs in MoFlow across nuScenes , NAVSIM v1 navtest , and NAVSIM v2 navhard .

Figure 4: Consecutive planning visualization on nuScenes. Comparison of MomAD ( Song et al., 2025 ) and MomWorld across consecutive frames. MomWorld yields more accurate and temporally consistent future trajectories.

Configuration	Value
Model input history	2 camera frames / 4 LiDAR sweeps
Image resolution	$2048\times 512$
Image backbone	V-99-eSE VoVNet
Planner latent / FFN width	256 / 1024
Planner attention layers / heads	3 / 8
Candidate vocabulary size	16,384

Table 9: NAVSIM implementation configuration of MomWorld.

Figure 5: Qualitative 6-second planning results on nuScenes. MomWorld generates smooth trajectories when transitioning from turns to straight driving, decelerating in dense traffic, turning left at intersections, and avoiding vehicles ahead.

Method	Venue	Closed-loop Performance				Multi-Ability (%) $\uparrow$
Method	Venue	DS $\uparrow$	SR (%) $\uparrow$	Effi. $\uparrow$	Comf. $\uparrow$	Merge	Overtake	Emergency Brake	Give Way	Traffic Sign	Mean
TCP-traj ∗ ( Wu et al., 2022 )	NeurIPS’22	59.90	30.00	76.54	18.08	12.50	22.73	52.72	40.00	46.63	34.92
UniAD ( Hu et al., 2023 )	CVPR’23	45.81	16.36	129.21	43.58	14.10	17.78	21.67	10.00	14.21	15.55
ThinkTwice ∗ ( Jia et al., 2023b )	CVPR’23	62.44	31.23	69.33	16.22	13.72	22.93	52.99	50.00	47.78	37.48
DriveAdapter ∗ ( Jia et al., 2023a )	ICCV’23	64.22	33.08	70.22	16.01	14.55	22.61	54.04	50.00	50.45	38.33
VAD ( Jiang et al., 2023 )	ICCV’23	42.35	15.00	157.94	46.01	8.11	24.44	18.64	20.00	19.15	18.07

Table 10: Closed-loop planning and multi-ability performance on Bench2Drive . ∗ denotes expert feature distillation. Higher is better for all metrics.

Figure 6: Qualitative planning results on NAVSIM. Comparison of GuideFlow ( Liu et al., 2026a ) and MomWorld across turning, straight driving, braking, stable driving, and vehicle avoidance. MomWorld produces smoother trajectories with improved roadway compliance and safer interactions.

Method	Turning-nuScenes					Adv-nuSc
	Avg. $L_{2}$ (m) $\downarrow$	Col. Rate (%) $\downarrow$				Col. Rate (%) $\downarrow$
	Avg. $L_{2}$ (m) $\downarrow$	1s	2s	3s	Avg.	1s	2s	3s	Avg.
UniAD	–	–	–	–	–	0.800	4.100	6.960	3.950
VAD	–	–	–	–	–	4.460	7.590	9.080	7.050
SparseDrive	0.86	0.04	0.17	0.98	0.40	0.029	0.618	2.430	1.026
DiffusionDrive ∗	–	0.03	0.14	0.85	0.34	0.068	1.299	3.646	1.671

Table 11: Open-loop robustness on Turning-nuScenes ( Song et al., 2025 ) and Adv-nuSc ( Xu et al., 2025 ) . Turning-nuScenes additionally reports average

L_{2}

where available. Lower is better.

Method	Snow			Rain			Fog
Method	1s	2s	3s	1s	2s	3s	1s	2s	3s
SparseDrive	0.13	0.27	0.50	0.11	0.27	0.55	0.14	0.36	0.58
DiffusionDrive ∗	0.09	0.24	0.39	0.07	0.18	0.35	0.06	0.18	0.30
MomAD	0.08	0.16	0.30	0.06	0.17	0.31	0.06	0.19	0.32
DIVER	0.07	0.13	0.25	0.05	0.16	0.27	0.04	0.16	0.25
GraphWorld	0.07	0.15	0.28	0.05	0.15	0.29	0.05	0.17	0.29

Table 12: Collision rate under three adverse-weather corruptions on nuScenes-C ( Dong et al., 2023 ) . Lower is better.

Explore similar work

MomADv2: Reliable Temporal Memory for End-to-End Autonomous Driving
Aug 24, 2026Ziying Song, Shengkai Zhang, Lin Liu +8Causality-Aware End-To-End Autonomous Driving Autonomous Driving
School of Artificial Intelligence (School of Software), Yanshan University · Beijing Jiaotong University · Nanyang Technological University +5
GraphWorld: Long-Horizon Planning with World Models for End-to-End Autonomous Driving
Jun 15, 2026Ziying Song, Caiyan Jia, Lin Liu +8Causality-Aware End-To-End Autonomous Driving Autonomous Driving
Beijing Key Laboratory of Traffic Data Mining and Embodied Intelligence, School of Computer Science and Technology, Beijing Jiaotong University, Beijing, China
DriveFuture: Future-Aware Latent World Models for Autonomous Driving
May 10, 2026Yufeng Hong, Xiaotian Zhou, Yingyan Li +6Autonomous Driving Latent World Models
Beijing Institute of Technology · Institute of Automation, Chinese Academy of Sciences · Beihang University +5

Aug 24, 2026, cs.CVJ/K move · Enter open · S save

MomADv2: Reliable Temporal Memory for End-to-End Autonomous Driving

Ziying Song, Shengkai Zhang, Lin Liu, Peiliang Wu+7

School of Artificial Intelligence (School of Software), Yanshan University · Beijing Jiaotong University · Nanyang Technological University+5

Causality-Aware End-To-End Autonomous Driving Autonomous Driving Long-Horizon Planning Long Refinement Trajectories

Open paper AskPDF arXiv

Long-horizon planning is critical for safe autonomous driving in complex scenarios. Existing methods improve planning continuity with temporal memory, but such memory may become invalid and mislead decisions when the driving command changes. Thus, selectively leveraging useful history while suppressing command-inconsistent memory remains a key challenge. To address this issue, we propose MomADv2, a reliable state-space memory framework for long-horizon end-to-end autonomous driving. At its core, MomADv2 introduces a Selective State-Space Planning Memory Query Module, which filters historical planning queries based on temporal continuity and command consistency, selects planning modes relevant to the current command, and models the evolution of planning intentions through a selective state-space mechanism. To further alleviate local trajectory deviations and error accumulation in long-horizon planning, we design a Flow-Matching Trajectory Residual Refiner. It learns a continuous residual correction field from the refined planning output to the expert trajectory, enabling fine-grained trajectory refinement while preserving the stability of anchor-based planning. Extensive experiments on closed-loop NAVSIM and Bench2Drive, as well as open-loop nuScenes, demonstrate that MomADv2 improves long-horizon planning consistency and reduces the average collision rate by 15.6% over MomAD under 6-second planning.

Abstract

Explore similar work

MomADv2: Reliable Temporal Memory for End-to-End Autonomous Driving

GraphWorld: Long-Horizon Planning with World Models for End-to-End Autonomous Driving

DriveFuture: Future-Aware Latent World Models for Autonomous Driving