Driving, Fast or Slow? Neuro-Symbolic Guidance for Motion Prediction in Multi-Modal Ground Mobility

arXiv:2606.15251 · cs.RO, cs.AI, cs.LG · Submitted 2026-06-13 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: Today's paper: "Driving, Fast or Slow? Neuro-Symbolic Guidance for Motion Prediction in Multi-Modal Ground Mobility".

Dev: The gist The proposed Trajectory Compliance-Shaping (TraCS) framework is a neuro-symbolic system that augments existing motion prediction backbones with interpretable and probabilistic first-order logic to align predictions with real-world traffic…

Rosa: First, who's behind it and why it matters.

Title and authors: Rosa: So we're looking at this paper today titled "Driving, Fast or Slow? Neuro-Symbolic Guidance for Motion Prediction in Multi-Modal Ground Mobility". It looks like they're tackling the problem of making motion predictions more accurate and interpretable in complex traffic scenes with different types of vehicles.

Dev: Yeah, it seems to be about taking those existing motion prediction models, which are often just black boxes, and adding some kind of structured reasoning on top to make them follow real-world traffic rules. It’s not just about guessing the next move; it’s about making sure that guess actually makes sense according to the rules of the road or whatever the local convention is.

Taro: I'm curious how they handle the fact that this has to work across different modes like cars, bikes, and pedestrians all at once. It sounds like a big challenge just getting those different behaviors to talk to each other correctly in a prediction system.

Rosa: Exactly, and the authors are proposing this Trajectory Compliance-Shaping framework as a way to do that integration. They’re using first-order logic to define what compliant behavior looks like in a given traffic context. It's trying to give the neural prediction backbone some explicit rules to follow instead of just relying on learned patterns from data, which is where the problem usually comes up when things get weird in real life.

Dev: That’s the core idea, right? Instead of just saying "this car usually goes there," they are trying to say "this car *must* slow down here because of this regulatory constraint." It moves the system from pure statistical prediction to something that incorporates explicit knowledge about how things should behave.

Taro: I wonder if that explicit knowledge is robust enough when the rules themselves are described in natural language, which is where they seem to be bringing in large language models to help generate those rules. That’s a new layer of complexity we haven't seen much of before in motion prediction research.

Rosa: That's what they did; they use an agentic code-generation pipeline that takes a natural language description of traffic regulations and turns it into these formal logic programs. It’s bridging that gap between human-readable rules and the math the prediction backbone understands.

The paper's summary: Dev: So, what does this TraCS framework actually propose? Essentially, they are taking an existing motion prediction backbone and plugging this TraCS module into it. This module uses that generated first-order logic to guide how the backbone weights its proposed trajectories.

Rosa: Right, so the main idea is this plug-and-play system that enriches the existing neural model with symbolic and probabilistic reasoning. They introduce a way to define spatial relations using a probabilistic vocabulary based on state transitions, which they call Transitional StaR Maps.

Taro: Transitional StaR Maps sound complicated. Can you break down what that means for us in terms of actual movement? Are we talking about defining things like "if the agent is at speed x, then it must follow rule r "?

Dev: Basically, they are creating a vocabulary for spatial relations over agent states and state transitions. This allows them to define logical predicates that can depend on things like agent speed or restrict certain directed movements. It formalizes how an agent's state changes into the next state in a way that’s mathematically structured.

Rosa: And then they use this vocabulary to calculate a compliance probability, which is P(C x,, delta t, R), which is the probability that the predicted path across all future states satisfies the specific compliance clause C defined in their rules R. They get this by assigning probabilities to logical atoms based on those StaR Maps and then doing a point-wise compliance score using weighted model counting over all models of the program.

Taro: So, they are essentially checking every possible future trajectory against the logic rules, giving it a score, and then fusing that with the original prediction. That sounds like a lot of computation for real-time use.

Dev: It is computationally intensive in its calculation method, but what they do afterward is crucial: they use this compliance probability to reweight the backbone’s predictions based on an estimated level of compliance and contextual alignment, which ensures graceful degradation if the input rules are imperfect or if there are non-compliant participants.

The paper's improvements: Rosa: One big improvement they highlight is how this system handles uncertainty through a learned neural confidence rating. They fit a backbone-specific trust network that outputs a score, say in the range of zero to one. This acts as a context-aware attenuation of the compliance signal.

Taro: That sounds like an important safety feature. So if the system is unsure about what the rules actually mean in this specific situation, it should dial down how much it relies on that symbolic logic, right?

Dev: Precisely. That neural confidence rating prevents overconfidence from dominating the decision-making process. It ensures that when the encoding is imperfect or there are non-compliant traffic participants, like maybe a cyclist breaking a rule, the system doesn't just blindly follow a flawed symbolic rule and cause a failure.

Rosa: And then they use this gated fusion for tail-risk treatment for the predictor. They selectively repair the worst-case scenarios that would dominate downstream planner risk by applying this symbolic reasoning only when TraCS is likely to be informative about compliance.

Taro: So it’s not always on, and when it is on, it’s specifically targeting those high-risk situations where a purely neural prediction might be dangerously wrong. That sounds like a smart way to apply the symbolic layer without slowing everything down constantly.

Dev: The analysis showed that this gated fusion actually improves prediction calibration across different mobility modes. For instance, infrequently observed classes like buses and cyclists, which tend to have left-skewed distributions, get better alignment. But for vehicles, which dominate most of the data and are narrowly peaked around zero with a near-balanced sign profile, the system correctly leaves predictions alone when TraCS doesn't offer much information gain.

Conclusion: Rosa: So to wrap up this paper on "Driving, Fast or Slow? Neuro-Symbolic Guidance for Motion Prediction in Multi-Modal Ground Mobility," the main point is that combining a neural prediction backbone with interpretable and probabilistic first-order logic gives us a way to explicitly encode real-world traffic constraints.

Dev: They achieved this by using an agentic code generation pipeline to turn natural language regulations into logic programs, which then guide the trajectory distribution through weighted model counting over all possible models of the program. It shows that symbolic reasoning is a computationally efficient complement to purely neural motion predictors when you need that level of compliance detail.

Taro: What I think stands out for me is how they tackle the uncertainty by using that learned neural confidence rating to gate the compliance signal, which prevents overconfidence and makes it robust when things get messy on the road.

Rosa: Yeah, it’s about making sure the system degrades gracefully instead of failing hard when it encounters something unexpected or non-compliant. It’s a way to treat those worst-case scenarios that would normally dominate downstream planning risk by only applying the symbolic logic where it actually adds value.

Dev: This work provides a framework for motion prediction in multimodal ground mobility environments, spanning cars, bicycles, and buses, showing that embedding structured compliance reasoning on top of neural representations yields critical gains at the tail end of the distribution.

Taro: It’s a solid step toward making autonomous systems more reliable by explicitly grounding them in traffic conventions rather than just statistical noise.

Rosa: That’s what this paper is about. We'll be right back after the break to look at some work on industrial multi-robot systems, and how they are using verification-gated agentic mission-state governance.

Simon Kohaut, Felix Divo, Julius Hahnewald, Benedict Flade, Julian Eggert, Kristian Kersting, Devendra Singh Dhami

Artificial Intelligence and Machine Learning Lab, TU Darmstadt · Honda Research Institute Europe (HRI-EU) · Hessian Center for AI (hessian.AI) · Centre for Cognitive Science · German Center for AI (DFKI) · Uncertainty in Artificial Intelligence Lab, TU Eindhoven

cs.RO, cs.AI, cs.LG

Submitted: 2026-06-13

Updated: 2026-10-14

Comments: Accepted at CoRL 2026 (Oral)

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 86/100

The gist: The gist The proposed Trajectory Compliance-Shaping (TraCS) framework is a neuro-symbolic system that augments existing motion prediction backbones with interpretable and probabilistic first-order

Key concepts

Trajectory Compliance-Shaping (TraCS)
A neuro-symbolic framework that augments motion prediction backbones with interpretable, probabilistic first-order logic. It integrates natural language traffic regulations to align the model's predicted trajectories with real-world mobility constraints, offering more transparent and safer predictions.
Agentic Code-Generation Pipeline
A system used to bridge the gap between human descriptions of traffic rules (natural language) and formal logical programs. It involves a coder, judge, and verifier tool that iteratively generates declarative first-order logic programs describing compliant behavior for different mobility modes.
Transitional StaR Maps
A probabilistic vocabulary used to formally define spatial relations over agent states and transitions in urban mobility. These maps allow the system to create logical predicates that depend on factors like agent speed or restrict certain movements based on context.

Terminology

Summary

The gist The proposed Trajectory Compliance-Shaping (TraCS) framework is a neuro-symbolic system that augments existing motion prediction backbones with interpretable and probabilistic first-order logic to align predictions with real-world traffic regulations, showing consistent improvements on state-of-the-art motion forecasting models

How it works

TraCS is designed as a plug-and-play module that integrates natural-language prompts about local traffic regulations to align a backbone’s weight distribution across its proposed trajectories with agreed-upon traffic conventions. To achieve this, TraCS employs an agentic code-generation pipeline to bridge the gap between natural-language descriptions of traffic regulations and probabilistic motion prediction. This pipeline generates declarative first-order logic programs that describe the compliant behavior of each mobility mode in a given traffic context.

Symbolic Reasoning and Modeling

The framework introduces a probabilistic, first-order vocabulary for spatial relations over agent states and state transitions. Specifically, it follows and extends the concept of Statistical Relational Maps (StaR Maps), which formalize reasoning over uncertain spatial relations in urban mobility. Transitional StaR Maps define transitional relations rδt(x, x˙, F) that allow for the definition of logical predicates that, e.g., depend on agent speed or restrict certain directed movements.

Trajectory Scoring and Fusion

TraCS determines the probability P(C x, x˙, δt, R), i.e., the probability that across τk each x and its transition into the next state satisfy the compliance clause C defined in the rules R. This is calculated by assigning probabilities to logical atoms using an estimated Transitional StaR Map and then obtaining a point-wise compliance score via weighted model counting over all models J of the program. The final fused probability is calculated using a log-linear opinion pool where the symbolic log-probability is centered against the log-uniform baseline log(1/6)

Trust and Confidence Gating

To prevent overconfidence, TraCS employs a neural confidence rating learned as a context-aware attenuation of the compliance signal. This is achieved by fitting a backbone-specific trust network that emits a trust score wˆ ∈ [0, 1] TraCS "consistently improves state-of-the-art prediction backbones, showing that probabilistic and symbolic compliance reasoning is a broadly applicable and computationally efficient complement to purely neural motion predictors. Furthermore, the system functions as a tail-risk treatment for the predictor, selectively repairing the worst-case scenarios that would dominate downstream planner risk".

Impact Across Mobility Modes

Analysis shows that TraCS improves prediction calibration across modes of mobility, with infrequently-observed classes, i.e., buses, cyclists, and pedestrians showing the most left-skewed distributions while for vehicles, which dominate the dataset, the distribution is narrowly peaked around zero with a near-balanced sign profile. This indicates that TraCS's gated fusion correctly leaves predictions intact where TraCS provides no information gain.

Conclusion

The paper concludes that TraCS is a neuro-symbolic framework for motion prediction in multimodal ground mobility environments, spanning agents such as cars, bicycles, and buses by embedding structured compliance reasoning on top of neural representations yield critical gains at the tail end.

Limitations and Future Work

A limitation noted is that the framework treat[s] the probabilistic logic program as an inference-time artifact, regenerated via LLM prompting rather than optimized during end-to-end training. A future direction suggested is to close this loop, using the probabilistic logic structure as a differentiable inductive bias through which neural sensor representations could be learned end-toend. Additionally, the model currently treats agents as conditionally independent given the scene context and does not explicitly represent joint maneuvers or negotiations between agents.

Acknowledgements

The work acknowledges financial support from Honda Research Institute Europe (HRI-EU) and other grants including those from the German Federal Ministry for Economic Affairs and Energy (BMWE) through “EU-SAI: Souveräne KI für Europa”.

References

The paper cites numerous related works, including foundational texts on probabilistic robotics like Probabilistic Robotics and various motion prediction models such as QCNet and HPNet. The paper also references work on neuro-symbolic systems such as DeepProbLog and NeurASP. The reference list includes works on datasets like Argoverse 2 and benchmarks like Waymo Motion Prediction Challenge. The paper also references specific papers related to StaR Maps and agentic code generation tools for ASP and LLM-based agent synthesis for code generation.

How it works

The paper details the Agent Encodings mapping between different components such as Regulations and Probabilistic First-Order Logic and how these map to the prediction backbone's output.

How it works

The paper describes the process of agentic codification in detail, involving a coder, a judge, and a verifier tool. The coder generates candidate Resin programs based on natural language descriptions and the judge evaluates them for faithfulness to the original intent. This iterative loop continues until both the alignment agent and the compiler are satisfied.

How it works

The scoring mechanism involves calculating P(C x, x˙, δt, R) using weighted model counting over the set of all models J of the program.

Improvements for AI systems

  1. Neuro-Symbolic Augmentation for Traffic Prediction: Integrate TraCS as a plug-and-play module that augments existing black-box motion prediction backbones with interpretable and probabilistic first-order logic. This allows existing neural models to be enhanced by explicitly encoding regulatory and behavioral constraints of real-world mobility, moving beyond purely statistical pattern matching.

  2. Agentic Code Generation for Rule Encoding: Utilize an agentic code-generation pipeline to bridge the gap between natural-language descriptions of traffic regulations and probabilistic motion prediction by having an LLM generate declarative first-order logic programs based on context, ensuring the symbolic reasoning is semantically grounded in the original regulations.

  3. Probabilistic Logic Layer for Spatial Reasoning: Employ a hybrid probabilistic atoms obtained by fitting statistical models to scene features to create a principled probabilistic logic layer that captures the uncertainty inherent in real-world environments, which serves as the vocabulary for expressing traffic regulations.

  4. Context-Aware Confidence Rating for Robust Steering: Implement a neural confidence rating learned as a context-aware attenuation of the compliance signal to prevent overconfidence, ensuring graceful degradation in the face of imperfect encodings, non-compliant traffic participants, or wrongful traffic rule prompts.

  5. Tail-Risk Mitigation via Gated Fusion: Apply a mechanism where gated TraCS fusion functions as a tail-risk treatment for the predictor, selectively repairing the worst-case scenarios that would dominate downstream planner risk by applying symbolic reasoning only when TraCS is likely to be informative.

Abstract

Accurate and interpretable motion prediction for heterogeneous traffic spaces, including pedestrians, bicycles, cars, and trucks, is essential for safe autonomous navigation. Nevertheless, state-of-the-art approaches remain predominantly black-box, lacking explicit encoding of the regulatory and behavioral constraints of real-world mobility. We propose Trajectory Compliance-Shaping (TraCS), a neuro-symbolic framework that augments existing black-box motion prediction backbones with interpretable and probabilistic first-order logic. To do so, TraCS employs an agentic code-generation pipeline to bridge the gap between natural-language descriptions of traffic regulations and probabilistic motion prediction. Furthermore, TraCS employs a reactive data-streaming inference engine that maintains and efficiently updates compliance landscapes as scenes evolve. To prevent TraCS from overconfidently steering the backbone's predictions in the wrong direction, we propose a neural confidence rating learned as a context-aware attenuation of the compliance signal. We demonstrate on the Argoverse 2 benchmark how TraCS consistently improves state-of-the-art prediction backbones, showing that probabilistic and symbolic compliance reasoning is a broadly applicable and computationally efficient complement to purely neural motion predictors.

Sources

Related papers