ManeuverFlow: Maneuver Hypothesis Generation with Future Evidence for Trajectory Evaluation and Refinement in End-to-End Autonomous Driving

What it claims

ManeuverFlow's main idea is to keep every candidate trajectory as an explicit, separate hypothesis from generation through to final selection. The paper contrasts this with MeanFuser, whose Adaptive Reconstruction Module fuses its one-step proposals into a single output. It also contrasts with WoTE, which runs a separate world-model rollout for every candidate. ManeuverFlow instead predicts one future BEV field per planning step that all candidates share. Path-specific evidence comes from a learned probe that samples that shared field along each trajectory, so no per-candidate forecasting is needed.

The headline numbers come from two different perception variants:

• NAVSIM: 90.58 PDMS (v1) and 90.14 EPDMS (v2) come from the ResNet-34 TransFuser (TF) variant, which uses cameras plus LiDAR.
• WorldEngine: 75.19 and 74.58 PDMS\* come from a camera-only Latent TransFuser (LTF) variant, which uses a learned latent in place of LiDAR.

How it works

1. Maneuver prior (M-GMP). Expert trajectories are partitioned into M motion families by displacement, mean heading and speed trend. A factorized Gaussian is fitted to each family in normalized diff-space. The authors' "maneuvers" are these trajectory-space clusters, not semantic labels. A scene-conditioned softmax predicts weights over the components.
2. Generator (M3F). This is a one-step MeanFlow generator built on the Improved Mean Flows (iMF) boundary-field and JVP construction. Its additions are the GMP sources and maneuver-routed feed-forward experts (the figure shows 8). Trajectory tokens cross-attend first to BEV/ego features, then to agent memories, with AdaLN time conditioning.
3. Candidate allocation. At inference, the top-3 maneuver components are kept and their weights renormalized. N candidates are then split among them by largest-remainder allocation.
4. Scoring (Metric Prism). Candidates attend to each other through set-level self-attention, then cross-attend to their own evidence. Heads predict five PDM factors: NC, DAC, EP, TTC and Comf. A fixed log-PDM composition ranks candidates on both NAVSIM v1 and v2. The v2-only factors are never used for selection.
5. Refinement (ECFR). The refiner predicts Frenet-frame residuals for R selected candidates. Its training target is the best evaluator-scored candidate with the same maneuver index (α_ref = 5, Smooth-L1 loss).

flowchart LR
  A[M-GMP top-3 components] --> B[M3F one-step bases]
  W[Shared future BEV field] --> P[Probe: path evidence]
  B --> P --> S[Metric Prism pre-score]
  S -->|R selected| R[ECFR Frenet refine]
  S -->|N−R retained| J[Joint re-score]
  R -->|re-query evidence| P
  R --> J --> O[Final trajectory]

Training runs in two stages:

• Stage 1 trains the auxiliary map, agent and maneuver losses together with the M3F loss.
• Stage 2 trains all modules jointly. At each step, the current model generates 64 candidates per navtrain scene. These candidates are labeled with the NAVSIM v1 evaluator to provide both the scorer's targets (weighted BCE) and the refiner's targets.

The future-BEV target comes from the same backbone, run with stop-gradient on future observations. That target is masked to the valid spatial overlap and supervised with Smooth-L1 plus cosine loss.

Evidence

• The only comparison figures in the fetched text are the headline scores above. The abstract also states that the WorldEngine score is the highest listed PDMS\*, under both logged and IDM-controlled traffic.
• A post-hoc analysis of the frozen M3F on 12,146 navtest scenes shows that best and top-κ PDMS improve as the candidate pool grows. Human driving scores 94.8 PDMS as a reference. This motivates refining a compact pool rather than enlarging it.

How to use it

Code: https://github.com/wyh-autonomy/maneuverflow. The page gives no install steps, weights or licence.

Re-implementing the training requires:

• the NAVSIM evaluator, used online during training to label 64 candidates per scene;
• future sensor frames, which the future-BEV target needs.

Limits

• The fetched text cuts off mid-method. No results tables, baseline numbers, ablations, latency or compute figures are available. The inference-time values of N and R are also not stated.
• The scorer is trained only on NAVSIM v1 factor labels. Some of the reported results use v2 metrics and the WorldEngine benchmark, which that training does not target.
• The NAVSIM and WorldEngine results come from different sensor configurations, so they are not one model's scores across benchmarks.