WAM-Flow: Parallel Coarse-to-Fine Motion Planning via Discrete Flow Matching for Autonomous Driving
WAM-Flow achieves parallel coarse-to-fine motion planning via discrete flow matching, reaching 89.1 PDMS with 1-step inference.
Key Findings
Methodology
WAM-Flow uses a vision-language-action model to cast ego-trajectory planning as discrete flow matching over a structured token space. The approach combines a metric-aligned numerical tokenizer, a geometry-aware flow objective, and simulator-guided GRPO alignment to enable parallel generation and coarse-to-fine inference.
Key Results
- On the NAVSIM v1 benchmark, WAM-Flow achieved 89.1 PDMS with 1-step inference and 90.3 PDMS with 5-step inference, significantly outperforming autoregressive and diffusion baselines.
- Compared to RecogDrive, 1-step inference speed increased by 4.67 times, while 5-step inference matched its latency.
- On the NAVSIM v2 benchmark, the full model achieved 84.7 EPDMS, demonstrating its robustness in complex scenarios.
Significance
This research introduces a new paradigm for end-to-end autonomous driving motion planning, addressing the efficiency bottleneck of autoregressive decoding and improving inference speed and accuracy through discrete flow matching. It holds significant implications for academia and industry, particularly in the development of efficient and safe autonomous driving systems.
Technical Contribution
WAM-Flow technically contributes by transforming autoregressive models into non-causal flow models, enhancing road-scene understanding through multimodal pretraining, and achieving higher closed-loop performance via parallel decoding inference.
Novelty
WAM-Flow is the first to apply discrete flow matching in autonomous driving, offering a novel motion planning method with significant efficiency and accuracy advantages over existing autoregressive and diffusion methods.
Limitations
- In complex urban environments, the model may require more computational resources to achieve high-precision trajectory planning.
- The model's performance depends on high-quality training data and may underperform in data-scarce scenarios.
Future Work
Future research could explore applying WAM-Flow in more complex driving scenarios and enhancing model robustness and accuracy by integrating more sensor data.
AI Executive Summary
The field of autonomous driving has long faced the challenge of achieving efficient motion planning while ensuring safety. Existing methods largely rely on autoregressive decoding, which is inefficient in complex scenarios. WAM-Flow introduces a new method of parallel coarse-to-fine motion planning through discrete flow matching.
At the core of WAM-Flow is the transformation of ego-trajectory planning into discrete flow matching over a structured token space. This method combines a metric-aligned numerical tokenizer, a geometry-aware flow objective, and simulator-guided GRPO alignment to enable parallel generation and coarse-to-fine inference. Experimental results show significant performance improvements on the NAVSIM v1 benchmark, achieving 89.1 PDMS with 1-step inference.
This research provides a new paradigm for end-to-end autonomous driving motion planning, addressing the efficiency bottleneck of autoregressive decoding and improving inference speed and accuracy through discrete flow matching. Future research could explore applying WAM-Flow in more complex driving scenarios and enhancing model robustness and accuracy by integrating more sensor data.
Deep Analysis
Background
The field of autonomous driving has seen significant advancements in recent years, yet motion planning remains a challenge. Traditional autoregressive methods are inefficient in handling complex scenarios, failing to meet real-time requirements. Recently, diffusion models and multimodal learning approaches have been introduced to the autonomous driving domain, but they still face trade-offs between inference speed and accuracy.
Core Problem
Motion planning in autonomous driving requires efficient trajectory prediction while ensuring safety. Existing methods largely rely on autoregressive decoding, which is inefficient in complex scenarios and fails to meet real-time requirements.
Innovation
WAM-Flow introduces a new method of parallel coarse-to-fine motion planning through discrete flow matching. This method combines a metric-aligned numerical tokenizer, a geometry-aware flow objective, and simulator-guided GRPO alignment to enable parallel generation and coarse-to-fine inference.
Methodology
- �� Use a metric-aligned numerical tokenizer to discretize continuous scalars into a shared codebook.
- �� Design a geometry-aware flow objective combined with simulator-guided GRPO alignment.
- �� Transform a pre-trained autoregressive model into a non-causal flow model through multi-stage adaptation.
Experiments
Experiments were conducted on the NAVSIM v1 and v2 benchmarks, using PDMS and EPDMS as primary evaluation metrics. The model achieved 89.1 PDMS with 1-step inference and 90.3 PDMS with 5-step inference, significantly outperforming autoregressive and diffusion baselines.
Results
WAM-Flow achieved 89.1 PDMS with 1-step inference on the NAVSIM v1 benchmark and 90.3 PDMS with 5-step inference, significantly outperforming existing autoregressive and diffusion methods.
Applications
WAM-Flow can be used for motion planning in autonomous driving systems, particularly in scenarios requiring high efficiency and safety. Its parallel inference capability provides significant advantages in complex scenarios.
Limitations & Outlook
In complex urban environments, the model may require more computational resources to achieve high-precision trajectory planning. Additionally, the model's performance depends on high-quality training data and may underperform in data-scarce scenarios.
Plain Language Accessible to non-experts
Imagine you're cooking in a kitchen. Traditional methods are like preparing each dish sequentially, chopping vegetables, then cooking, and finally serving. WAM-Flow is like preparing multiple dishes simultaneously, with several pots on the stove at once, saving time. Through this parallel approach, WAM-Flow can quickly adjust the taste and heat of each dish, ensuring each one reaches its optimal state. In autonomous driving, this method is like quickly planning the best driving route in a complex traffic environment.
ELI14 Explained like you're 14
Imagine you're playing a racing game. Traditional methods are like controlling one direction key at a time, slowly adjusting your direction. WAM-Flow is like pressing multiple direction keys at once, quickly adjusting your car's direction and speed. This method lets you dodge obstacles faster and find the best racing line in the game. Isn't that cool?
Glossary
Discrete Flow Matching
A method for transporting probability distributions from a simple distribution to a data distribution using a continuous-time Markov chain.
Used for motion planning in WAM-Flow.
Metric-Aligned Numerical Tokenizer
Discretizes continuous scalars into a shared codebook and maintains scalar geometry through triplet-margin learning.
Used for numerical tokenization in WAM-Flow.
Simulator-Guided GRPO Alignment
An alignment method integrating safety, progress, and comfort rewards while retaining parallel generation capabilities.
Used for closed-loop control in WAM-Flow.
PDMS
Predictive Driver Model Score, a composite measure evaluating the safety and performance of autonomous driving systems.
Used to evaluate WAM-Flow's experimental results.
EPDMS
Extended Predictive Driver Model Score, incorporating more sub-metrics for a comprehensive assessment of driving performance.
Used to evaluate WAM-Flow's performance on the NAVSIM v2 benchmark.
Open Questions Unanswered questions from this research
- 1 How to improve WAM-Flow's performance in data-scarce scenarios? Current methods rely on high-quality training data, and future work needs to explore data augmentation and transfer learning techniques.
Applications
Immediate Applications
Urban Autonomous Driving
WAM-Flow can be used in urban environments for autonomous driving, especially in complex traffic scenarios requiring rapid response.
Long-term Vision
Fully Autonomous Driving
With technological advancements, WAM-Flow is expected to achieve higher safety and efficiency in fully autonomous driving, becoming a core technology for future autonomous systems.
Abstract
We introduce WAM-Flow, a vision-language-action (VLA) model that casts ego-trajectory planning as discrete flow matching over a structured token space. In contrast to autoregressive decoders, WAM-Flow performs fully parallel, bidirectional denoising, enabling coarse-to-fine refinement with a tunable compute-accuracy trade-off. Specifically, the approach combines a metric-aligned numerical tokenizer that preserves scalar geometry via triplet-margin learning, a geometry-aware flow objective and a simulator-guided GRPO alignment that integrates safety, ego progress, and comfort rewards while retaining parallel generation. A multi-stage adaptation converts a pre-trained auto-regressive backbone (Janus-1.5B) from causal decoding to non-causal flow model and strengthens road-scene competence through continued multimodal pretraining. Thanks to the inherent nature of consistency model training and parallel decoding inference, WAM-Flow achieves superior closed-loop performance against autoregressive and diffusion-based VLA baselines, with 1-step inference attaining 89.1 PDMS and 5-step inference reaching 90.3 PDMS on NAVSIM v1 benchmark. These results establish discrete flow matching as a new promising paradigm for end-to-end autonomous driving. The code will be publicly available soon.