FlowDrive: moderated flow matching with data balancing for trajectory planning
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.
Rosa: We're looking at the paper: "FlowDrive: moderated flow matching with data balancing for trajectory planning".
Dev: Here's the gist of the paper. Learning-based planners are sensitive to the long-tailed distribution of driving data while FlowDrive achieves state-of-the-art results among learning-based planners and approaches methods with rule-based refinements.
Rosa: Problem Identification Trajectory planning in autonomous driving requires both safety and efficiency, but real driving data exhibits long-tailed distributions where common behaviors like lane-following dominate while rare but safety-critical cases are underrepresented.
Dev: This imbalance biases planners toward frequent patterns, reducing reliability in corner cases.
Rosa: Furthermore, dataset bias and limited diversity lead to poor generalization—especially in dynamic traffic scenarios unseen during training.
Title and authors: Rosa: So we're looking at this paper called FlowDrive: moderated flow matching with data balancing for trajectory planning. It’s tackling the problem that learning-based planners struggle because the real driving data is long-tailed, meaning common driving stuff dominates, and the really dangerous or rare stuff just isn't represented enough in the training sets.
Dev: Exactly. The imbalance biases these models toward frequent patterns, which makes them unreliable when you run into those critical edge cases that happen rarely but need to be handled safely. This paper claims they found a way to fix that using two main strategies: balancing how they sample the training data and then introducing a specific guidance mechanism for the planning process itself.
Taro: I’m interested in the idea of balancing by trajectory pattern, because if you can weight those rare motion patterns up, it should give the model much better coverage than just relying on general scenario labels. What does that actually look like in practice for someone trying to build a dataset?
Rosa: It means they use two ways to sample data. First is scenario-based sampling where they use inverse frequency weights to pick underrepresented types of driving scenarios. But the paper says the real improvement comes from cluster-based sampling, which precomputes clusters of normalized trajectories over the planning horizon and upweights samples from those rare clusters.
Dev: That cluster-based approach sounds like a solid way to flatten out that distribution, increasing exposure to a wider variety of motion patterns than just using simple scenario weights. Then they introduce FlowDrive itself, which is this flow-matching planner that learns a continuous motion flow to map noise directly into trajectory distributions with only a few steps.
Taro: So it’s not just sampling the input better; they're also changing the way the planning happens inside the model. How does this flow matching actually work in terms of computation? Is it fast enough for real-time control loops?
Rosa: They represent the trajectory as a vector in D dimensions and learn a time-dependent velocity field, v theta t x c, that moves a simple base distribution p zero to the actual conditional data distribution with just an ordinary differential equation integration. This lets them generate feasible driving trajectories very quickly because they only need a small number of flow steps for sampling.
Title and authors: Dev: That small number of flow steps is key for speed, but generating the trajectory is one thing, controlling it reliably in real-time is another. What about the safety aspect? How do you make sure this learned flow doesn't generate something completely nonsensical when the car hits a tricky spot?
Taro: That brings us to their second major contribution: moderated guidance. They inject small, structured perturbations directly inside that ODE integration to systematically increase trajectory diversity while still keeping things consistent with what's happening in the scene.
Rosa: It’s like they are nudging the planning process along a path of variation by applying an update, p prime pos t h equals p pos t h plus alpha of t beta of h (delta lat n hat plus delta lon tau hat), where that whole thing runs from one step up to H steps over the horizon.
Dev: That mechanism is designed to steer FlowDrive’s outputs toward more diverse trajectory samples, which the authors say leads to state-of-the-art performance on benchmarks like nuPlan and interPlan, even outperforming previous rule-based and hybrid planners. The paper shows that injecting this guidance at a specific point, Tg two equals one over two, gives the best trade-off between getting more diversity and keeping the generated paths feasible.
Taro: So what does that mean for a driver who just wants to know what’s actually different about this new system? It sounds like it’s unlocking behaviors that were just too rare in the original training data to ever be considered by the planner before.
Rosa: Right, because they found that this guided approach successfully unlocks multi-modal behaviors that were otherwise rare in their training data, and they show examples of this on things like overtaking when the context allows or maintaining speed limits during those overtaking maneuvers.
Dev: It’s a significant finding because it means the system isn't just repeating what it saw frequently; it’s learning how to explore possibilities that were previously ignored, which is crucial for real-world driving where you have to be ready for the unexpected.
Taro: I wonder about the limitations they mention in their own work. They note that they are still working on developing smoother constraints during the flow matching training and exploring richer learned guidance signals that encode things like safety or intent, steering with control vectors, or even reinforcement learning fine-tuning with methods like Flow Matching Policy Gradients.
Title and authors: Rosa: So while FlowDrive is state-of-the-art right now on nuPlan and interPlan, they acknowledge that the raw trajectories it generates are still jerky and they plan to work on improving the smoothness of those trajectories in future work.
Dev: And from a real-time control standpoint, you see that using two k flow steps for training and eight flow steps for inference gives them the best results, which shows how efficient this whole process is when you're looking at speed and latency.
Taro: It’s interesting to see that even with the complexity of trajectory planning, they found that generating just thirty candidates yields a score of zero point nine two nine six on nuPlan Val fourteen R, which is pretty close to their maximum performance score.
Rosa: And it’s important to remember the overall picture here: FlowDrive: moderated flow matching with data balancing for trajectory planning tackles the long-tailed distribution issue by combining smarter data sampling, a fast flow-matching mechanism, and targeted in-the-loop guidance. It shows how these components work together to move from just mimicking common driving patterns to actually exploring diverse and safe maneuvers.
Dev: So it’s about making learning planners robust enough for the messy reality of the road, not just the clean data they were fed. The implications are that we might see a step forward in autonomous systems handling complex, rare driving situations more gracefully than before.
Taro: For someone just listening to this, it means that when you think about how an autonomous car handles a sudden change in traffic or a tricky merge at an intersection, the underlying planning system is getting smarter at considering those less common ways things can go.
Rosa: That’s the core idea. We've seen this paper explore how to move beyond just following frequent patterns to actually generating diverse, yet feasible, driving trajectories by using these tailored data balancing and guidance techniques.
Dev: It’s a solid piece of research because it doesn't just throw more data at a model; it shows exactly how you can structure the training process and the generation process to get better results with less computational overhead during the actual planning step.
Taro: I just hope they keep working on those smoothness constraints, because jerky trajectories are a big deal when you’re talking about actual vehicle control and safety.
Rosa: Exactly, that's where the next iteration of this work will probably focus, moving from generating a good plan to generating a smooth, safe plan. That’s what we have here with FlowDrive: moderated flow matching with data balancing for trajectory planning.
The paper's summary: Rosa: So, to recap, FlowDrive is this planner that tackles the problem of long-tailed driving data by using two main techniques: smarter ways to pick your training examples and a specific way to guide the planning process itself.
Dev: Right, it’s about making sure the model sees those rare driving situations more often through these data balancing strategies, and then adding a mechanism that steers the generated paths toward more diverse options without breaking scene consistency.
Rosa: That steering mechanism is what they call moderated guidance, where they inject small nudges into the flow integration inside the ordinary differential equation solver to systematically increase trajectory diversity while keeping things looking like a real drive.
Dev: And this whole approach aims to push performance on benchmarks like nuPlan and interPlan to a new level by unlocking behaviors that were just too rare in their training data before.
Rosa: It means we're talking about getting the system to actually try out maneuvers it wasn't explicitly taught, like overtaking when the situation allows, which is pretty telling for real-world autonomy.
Dev: But there’s a trade-off they found—injecting that guidance at a specific point, one half way through the planning horizon, gives them the best balance between getting more variety and keeping those generated paths feasible.
Rosa: That’s interesting because it shows that just having a better data set isn't enough; you also need to be smart about how you use the planning model once it’s trained.
Dev: And from an engineering standpoint, I like that they found using eight flow steps for inference gives them the best result, showing how efficient this whole process is when you’re looking at speed and latency during actual operation.
Rosa: But they still have some work to do; they mentioned that the raw trajectories FlowDrive generates can still be a bit jerky, so future work will focus on making those paths smoother during the flow matching training itself.
Dev: True, and they also plan to explore ways to make those guidance signals even richer, maybe encoding things like safety or intent directly into the steering commands or fine-tuning the whole thing with reinforcement learning methods.
Rosa: So while FlowDrive is hitting state-of-the-art scores now, the next big challenge is taking those plans and making them actually look smooth and reliable when a car is actually driving it.
Dev: Exactly, so we’ve seen how to get high performance by combining better data sampling with a guided flow matching process, but the real test for this technology will be how smoothly it handles the messy, unpredictable reality of the road.
The paper's improvements: Taro: So, to recap, they’re talking about how FlowDrive can be improved by focusing on two main areas: making the training data better and refining how the planning process itself is guided during generation.
Rosa: Yeah, it means they found that balancing the training data using cluster-based sampling actually gives a bigger jump in performance than just using scenario labels alone, because it really helps flatten out those rare motion patterns.
Dev: And on top of that, the moderated guidance mechanism isn't just for diversity; it’s about injecting small nudges inside that ODE integration to ensure the trajectory stays consistent with what you actually see around you in a real scene.
Taro: That steering mechanism is key because it unlocks behaviors that were previously too rare in the training data to even be considered by the original planner.
Rosa: It suggests that if we want autonomous systems to handle those complex, unusual driving situations gracefully, we need to stop just feeding them common data and start actively guiding their exploration of possibilities.
Dev: From a controls angle, it’s about making sure the system doesn't just hallucinate a weird path; it's using that guidance to steer the generation toward paths that are still physically feasible within real-time constraints.
Taro: I wonder what happens when you push those boundaries—if this guidance can encode safety or intent, like telling the planner *why* it’s doing something, not just *what* to do.
Rosa: Exactly, they hint at the future by saying they want to look into richer guidance signals that actually convey safety information or what the vehicle's goal is when making those steering adjustments.
Dev: And we also saw some efficiency wins; using two thousand flow steps for training and only eight steps for inference gave them a good result, which shows this whole method can be fast enough to run in real-time loops.
Taro: So the next step isn't just about getting a higher score on a benchmark; it’s about making sure those diverse plans are also smooth and safe enough to execute without shaking the vehicle.
Rosa: Right, so FlowDrive is showing us how to get better results by balancing data and adding smart guidance, but the real challenge moving forward is taking that generated plan and making it actually drivable.
Conclusion: Rosa: So we've covered how FlowDrive uses data balancing to improve training coverage and moderated guidance to inject diversity into trajectory generation for autonomous driving.
Dev: That’s right, it shows how structuring the training data and adding that targeted steering mechanism can push planners past just mimicking frequent patterns.
Rosa: The implication is that learning-based systems can become much more robust when they are trained with this kind of thoughtful data coverage instead of just throwing in a lot of random driving examples.
Dev: It changes how we think about safety cases; if the planner can explore those rare scenarios better, we can build a more reliable safety case for real-world deployment.
Taro: I just see it as moving past the "happy path" planning where the car just follows the rules, toward a system that can actually handle unexpected stuff without breaking down.
Rosa: Right, and they’ve shown that even with these improvements, there's still room to grow—they’re looking at making those raw trajectories smoother because jerky movement is still a problem for actual vehicle control.
Dev: Exactly, and they want to explore more advanced guidance signals that tell the system not just what to do, but *why* it should do it based on safety or intent.
Taro: That’s where we need to go next; if the planner can understand intent, then the control systems built on top of it will be much better at handling complex interactions in crowded spaces.
Rosa: So that’s the story with FlowDrive: a strong start using data balancing and flow matching, with clear paths for making those plans smoother and smarter for real-world deployment.
Dev: It’s a solid piece of work that lays a good foundation for how we can train these models to be more dependable in messy situations.
Taro: We'll keep watching how they evolve this system into something that truly understands the nuances of driving, not just the basics.
Karlsruhe Institute of Technology
cs.RO, cs.AI, cs.LG
Submitted: 2025-09-26
Updated: 2026-11-19
Code: https://github.com/einsteinguang/flow_drive_planner
License: http://creativecommons.org/licenses/by-sa/4.0/
Importance score: 92/100
The gist: Learning-based planners are sensitive to the long-tailed distribution of driving data while FlowDrive achieves state-of-the-art results among learning-based planners and approaches methods with…
Key concepts
- Flow Matching Mechanism
- This is a method that learns a continuous motion flow. It maps random noise directly into desired trajectory distributions by solving an ordinary differential equation (ODE). This allows the planner to generate smooth, realistic trajectories quickly with very few steps.
- Data Balancing Strategy
- Since driving data often lacks rare events, this strategy uses two methods: scenario-based sampling and cluster-based sampling. Cluster-based sampling groups similar trajectories together and upweights rare clusters to ensure the planner is exposed to a wider variety of motion patterns.
- Moderated Guidance Mechanism
- This mechanism injects small, structured perturbations into the flow integration process. By applying updates based on lateral ($\delta_{lat}$) and longitudinal ($\delta_{lon}$) magnitudes, it systematically increases trajectory diversity while ensuring the generated paths remain consistent with the current scene.
Terminology
Summary
Learning-based planners are sensitive to the long-tailed distribution of driving data while FlowDrive achieves state-of-the-art results among learning-based planners and approaches methods with rule-based refinements. p. 0 We further introduce moderated, in-the-loop guidance that injects small perturbation between flow steps
to systematically increase trajectory diversity while remaining scene-consistent. On nuPlan and the interaction-focused interPlan benchmarks, FlowDrive achieves
state-of-the-art results among learning-based planners and approaches methods
with rule-based refinements.Where the paper says it
Problem Identification
Trajectory planning in autonomous driving requires both safety and efficiency, but real driving data exhibits long-tailed distributions where common behaviors like lane-following dominate while rare but safety-critical cases are underrepresented. This imbalance biases planners toward frequent patterns, reducing reliability in corner cases. p. 0 Real driving data exhibits long-tailed distributions, where
common behaviors like lane-following dominate while rare but safety-critical cases are underrepresented (Karnchanachari et al. This imbalance biases planners toward frequent patterns,
reducing reliability in corner cases. This imbalance biases planners toward frequent patterns,
reducing reliability in corner cases. Moreover, dataset bias and limited diversity lead to poor generalization—especially in dynamic traffic scenarios unseen during training (Codevilla et al.Where the paper says it
Where the paper says it
Flow Matching Mechanism
FlowDrive is a flow-matching planner that learns a continuous motion flow to map noise directly to trajectory distributions with few flow-matching steps, enabling fast, few-step sampling. The model represents the trajectory as a vector in R D and learns a time-dependent velocity field vθ(t, x, c) that transports a simple base distribution p0 to the conditional data distribution pdata(x c) by integrating an ordinary differential equation (ODE). p. 2 FlowDrive learns a time-dependent velocity field vθ(t, x, c) that transports a simple base distribution
p0 (e. , Gaussian over trajectory dimensions) to the conditional data distribution pdata(x | c) by
integrating an ordinary differential equation (ODE). , 2023) trains a parametric vector field vθ(t, x, c) to match
the target velocity that makes Equation 2 hold along the chosen path. With the rectified path, the
standard conditional RF/FM objective samples t ∼ U[0, 1], z ∼ p0, forms xt via Equation 1, and
minimizes
LRF(θ) = E(x,c)∼pdata(x,c),
z∼p0, t∼U[0,1]
h
w(t)
vθ(t, xt, c) − (x − z)
2
2
i
, (4)
where w(t) is an optional time-weighting schedule.Where the paper says it
Where the paper says it
Data Balancing Strategy
The paper introduces a data balancing scheme to mitigate long-tail biases in driving datasets, adopting two complementary sampling strategies. These include scenario-based sampling, where inverse-frequency weights are used to sample underrepresented scenario types, and cluster-based sampling, which precomputes clusters of normalized trajectories over the horizon and upweights samples from rare clusters to encourage exposure to a wider variety of motion patterns. p. 5 Given the cluster assignment for each training sample, we
compute inverse-frequency weights per cluster, analogous to the scenario-based scheme, and use
them in the weighted sampler. This encourages exposure to a wider variety of motion patterns (e.Where the paper says it
In-the-Loop Guidance
FlowDrive introduces a moderated guidance mechanism that injects small, structured perturbation inside the flow integration to systematically increase trajectory diversity while remaining scene-consistent. This involves defining lateral and longitudinal magnitudes δlat and δlon, and applying an update p'pos t,h = p pos t,h + α(t) β(h) (δlat nˆ + δlon τˆ), h = 1,..., H. The authors empirically found that injecting guidance at Tg 2 = 1/2 offers the best trade-off between diversity and feasibility. p. 6 Empirically, we found that injecting at Tg,2 = {
1
2
} offers the best trade-off between diversity and feasibility;
this choice is used by default in our experiments. In contrast to manual post-processing, our guidance approach keeps the model “in the loop”, letting
scene-conditioning absorb and correct the perturbation.Where the paper says it
Performance Results
FlowDrive achieves state-of-the-art performance on the nuPlan and interPlan benchmarks, outperforming previous rule-based, learning-based and hybrid baselines. The hybrid variant FlowDrive with cluster-based sampling (FlowDrive) surpasses previous learning-based planners by noticeable margins. p. 7 Without data balancing, FlowDrive− still outperforms
previous learning-based planners in almost all columns. With cluster-based sampling, FlowDrive
surpasses them by noticeable margins. On nuPlan and interPlan, FlowDrive outperforms all previous learning-based methods by large margins. With moderated guidance and light post-processing
(FlowDrive*), it achieves new state-of-the-art across nearly all benchmark splits while remaining
efficient. We compare the three sampling strategies on nuPlan
Val14 (R). As shown in Table 2a, scenario-label imbalance is not the main bottleneck: reweighting by scenario type performs worse than no weighting (FlowDrive−), while balancing by trajectory
pattern (FlowDrive) yields the largest improvement. , 2024) and the
interaction-focused interPlan benchmark (Hallgarten et al. On nuPlan, we report results on
the official Val14, Test14, and Test14-hard splits in both non-reactive and reactive simulator modes. Remarkably, despite using no post-processing, its scores are
very close to the strong hybrid baselines PDM-Hybrid and PLUTO. Adding moderated guidance and
post-processing (FlowDrive*) then pushes performance further to the top, achieving state-of-the-art
across nearly all columns among all planner categories. , 2024) and the
interaction-focused interPlan benchmark (Hallgarten et al. On nuPlan, we report results on
the official Val14, Test14, and Test14-hard splits in both non-reactive and reactive simulator modes.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Future Directions
Future work will investigate the jerky raw trajectories from FlowDrive and try to develop improved smoothness constraints during flow matching training. The authors plan to further explore richer learned guidance signals encoding safety or intent, steering with control vectors, and RL fine-tuning with Flow Matching Policy Gradients or ReinFlow. p. 8 We will further explore richer learned guidance signals encoding safety or intent (Feng et al. , 2025), steering with control vectors (Tas & Wagner, 2025) and RL fine-tuning with Flow Matching Policy Gradients or ReinFlow (McAllister et al.Where the paper says it
Ablation Insights
Ablation studies on data balancing show that balancing by trajectory pattern yields the largest improvement in performance compared to scenario-label imbalance. Ablations on training and inference steps indicate that using 2k flow steps at training and 8 flow steps at inference gives the best result. p. 8 09
sampling method, and apply different number of flow steps at inference (Table 2b). We find that
using 2k flow steps at training and 8 flow steps at inference gives the best result, which achieves
runtime of 40 milliseconds per inference pass on a single NVIDIA GeForce RTX 2080 Ti GPU. ,
2023), and executing the top-scoring plan. We provide ablation study of the number of candidates
on driving scores in Section A. With cluster-based sampling, FlowDrive
surpasses them by noticeable margins. Remarkably, despite using no post-processing, its scores are
very close to the strong hybrid baselines PDM-Hybrid and PLUTO. InterPlan is a closed-loop driving benchmark based on the nuPlan dataset and framework. It modifies
original nuPlan scenarios by augmenting agent counts and behaviors so that more intelligent maneuvers become necessary, such as sudden intruding pedestrians and nudging around parked vehicles. PlanTF: a transformer planner that formulates planning as sequence modeling with endto-end training on expert demonstrations (Cheng et al. PLUTO: an imitation-learning planner enhanced with contrastive learning and augmentations that achieves strong closed-loop scores
on nuPlan (Cheng et al.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Qualitative Demonstrations
FlowDrive demonstrates robust performance across diverse driving situations, including waiting for pedestrians, intersection navigation, car-following, and complex maneuvers like nudging around parked vehicles and emergency stopping. FlowDrive successfully executes the lane change maneuver in an interPlan scenario where previous versions failed. p. 1 • We introduce a guidance mechanism that steers FlowDrive’s outputs to produce more diverse trajectory samples. This enables state-of-the-art performance on the nuPlan and interPlan benchmarks, outperforming previous rule-based, learning-based and hybrid baselines. Lateral offsets often
unlock multi-modal behaviors that are otherwise rare in
training data. For instance, in car-following with a slow
leading vehicle, small offsets encourage overtake-like
solutions if the adjacent lane is free; see Figure 5. Diversity and Multi-modality. Lateral offsets often
unlock multi-modal behaviors that are otherwise rare in
training data. This in-the-loop perturbation lets the learned velocity field subsequently
reconcile the guided displacement with scene context, unlike post-hoc shifts. Lateral offsets (δlat)
promote modal diversity (overtaking, nudging), while longitudinal offsets (δlon) can reshape speed
profiles.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Training Details
The model is trained on 1M sampled scenarios from nuPlan dataset using 8 NVIDIA RTX 6000 Ada GPUs over about 40 hours for 400 epochs. The encoder and decoder are wrapped in an exponential moving average (EMA) model. p. 3 In FlowDrive, x represents the ego-trajectory, c encodes scene context, and vθ is parameterized by
a context encoder and planning decoder. Next, we explain the model architecture of FlowDrive in
Section 3. Ablation on Flow Training and Inference Steps. We use Euler ODE solver to discretize flow
time steps. We further introduce moderated, in-the-loop guidance that injects small perturbation between flow steps
to systematically increase trajectory diversity while remaining scene-consistent. On nuPlan and the interaction-focused interPlan benchmarks, FlowDrive achieves
state-of-the-art results among learning-based planners and approaches methods
with rule-based refinements. After adding moderated guidance and light postprocessing (FlowDrive*), it achieves overall state-of-the-art performance across
nearly all benchmark splits. Our code is available at https://github. 0
2
4
6
8
10
12
14
16
18
Cluster ID
102
103
104
Sample Count (log scale)
No weighted sampling
0
2
4
6
8
10
12
14
16
18
Cluster ID
Scenario-based sampling
0
2
4
6
8
10
12
14
16
18
Cluster ID
Cluster-based sampling
Cluster Distribution Per Sampling Method (n batches=100, batch size=100)
Figure 3: Sample counts per cluster after applying different training-time sampling strategies. Cluster-based sampling produces the most balanced distributions over the clusters. We follow a scene input representation similar to Diffusion
Planner (Zheng et al. First, all coordinates are in the local ego frame, with the origin at the
current ego position and heading aligned with the ego vehicle’s heading direction. For a batch of
size B: neighbor history has shape [B, Pn, Tp, Fn] with positions, kinematics, and a one-hot agent
type; static objects are [B, Ps, Fs]; lanes are polylines[B, Pℓ, V, Fℓ] with local geometry, traffic-light
signals, route membership and per-lane speed limit features. We denote the full context as c, and the
trajectory sample as a sequence of H steps with action dimension A (A = 4 for x, y, cos,sin); we
flatten x ∈ R
H×A to R
D when referring to Equation 4.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Data balancing strategy for training improves robustness by increasing coverage of rare behaviors. Specifically,
Cluster-based sampling substantially flattens the cluster distribution and increases coverage of rare motion patterns
compared to no weighting or scenario-based sampling. -
FlowDrive generates feasible driving trajectories efficiently through a learned rectified flow, which
minimizes LRf(theta) = E(x,c)∼pdata(x,c), z∼p0, t∼U[0,1] h w(t) vθ(t, xt= (1 − t) z + t x
and allows for fast sampling via asmall number of flow steps.
-
The moderated guidance mechanism steers generated trajectories to increase diversity while maintaining scene consistency by injecting small perturbations inside the ODE integration. This is achieved by applying an update
p′pos t,h = p pos t,h + α(t) β(h) (δlat nˆ + δlon τˆ), h = 1,..., H
before evaluating the next velocity field in the ODE solver. -
The improved system FlowDrive with moderated guidance and light post-processing achieves state-of-the-art performance by
unlock[ing] multi-modal behaviors that are otherwise rare in training data,
demonstrated by successfully executing maneuvers likeovertaking when context allows
and maintaining proper speed limits during overtaking scenarios.
Abstract
Learning-based planners are sensitive to the long-tailed distribution of driving data. Common maneuvers dominate datasets, while dangerous or rare scenarios are sparse. This imbalance can bias models toward the frequent cases and degrade performance on critical scenarios. To tackle this problem, we compare balancing strategies for sampling training data and find reweighting by trajectory pattern an effective approach. We then present FlowDrive, a flow-matching trajectory planner that learns a conditional rectified flow to map noise directly to trajectory distributions with few flow-matching steps. We further introduce moderated, in-the-loop guidance that injects small perturbation between flow steps to systematically increase trajectory diversity while remaining scene-consistent. On nuPlan and the interaction-focused interPlan benchmarks, FlowDrive achieves state-of-the-art results among learning-based planners and approaches methods with rule-based refinements. After adding moderated guidance and light post-processing (FlowDrive*), it achieves overall state-of-the-art performance across nearly all benchmark splits. Our code is available at https://github.com/einsteinguang/flow drive planner.
Sources
- Rethinking Imitation-based Planner for Autonomous Driving
- PLUTO: Pushing the Limit of Imitation Learning-based Planning for Autonomous Driving
- On the Guidance of Flow Matching
- Mastering Diverse Domains through World Models
- Can Vehicle Motion Planning Generalize to Realistic Long-tail Scenarios?
- Classifier-Free Diffusion Guidance
- GameFormer: Game-theoretic Modeling and Learning of Transformer-based Interactive Prediction and Planning for Autonomous Driving
- Towards learning-based planning:The nuPlan benchmark for real-world autonomous driving
- Improving the Training of Rectified Flows
- DiffusionDrive: Truncated Diffusion Model for End-to-End Autonomous Driving
- Flow Matching for Generative Modeling
- Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow
- Exponential Moving Average of Weights in Deep Learning: Dynamics and Benefits
- FlowMP: Learning Motion Fields for Robot Planning with Conditional Flow Matching
- Towards Balanced Behavior Cloning from Imbalanced Datasets
- Plan-R1: Safe and Feasible Trajectory Planning as Language Modeling
- Words in Motion: Extracting Interpretable Control Vectors for Motion Transformers
- Enhancing Autonomous Driving Safety with Collision Scenario Integration
- Quantum optimal transport with convex regularization
- Diffusion-ES: Gradient-free Planning with Diffusion for Autonomous Driving and Zero-Shot Instruction Following
Related papers
- Rethinking Actions: Continuity-Aware Priors for World-Model Planning
- MOSAIC: Modular Obstacle and Spatial-Aware Framework for Manipulation In Cluttered Scenes
- SAGAS: Semantic-Aware Graph-Assisted Stitching for Offline Temporal Logic Planning
- DexSeed: Expanding Human Hand-Object Interactions into Robot Demonstrations for Dexterous Policy Learning
- LaST0.5: A Generalist World Action Model with Interaction-Centric Latent Reasoning
- CIALLO: Learning Shared Intent Representations for Cross-Embodiment Manipulation