Constraint Enforcement in Generative Trajectory Planning

31 concepts28 papers1 proposal

Generative models, particularly diffusion models and flow matching, have revolutionized how robots and autonomous vehicles plan their movements. By learning from vast datasets of expert demonstrations, these models can generate diverse, smooth, and natural-looking trajectories that capture the complexity of real-world behavior. However, a fundamental "safety gap" remains: while these models are excellent at statistical imitation, they are often "blind" to the hard physical and legal constraints of the environments they inhabit.

As noted by Zifan Zeng et al., even high-fidelity world models can hallucinate physically impossible scenarios or violate basic traffic rules. Because these models prioritize the general shape of the training data over strict boundaries, they often "color outside the lines," leading to collisions, off-road drifting, or infeasible maneuvers.

The Safety Gap in Generative Models

The core problem lies in the probabilistic nature of generative planning. A standard diffusion model learns the distribution of expert trajectories, but it does not inherently understand why a trajectory is safe. Anjian Li et al. point out that small prediction errors in these models can accumulate, causing generated paths to violate geometric constraints like clipping walls or intersecting obstacles.

This lack of inherent constraint satisfaction manifests in several ways:

  • Trajectory Divergence: Predicted paths may drift away from road boundaries or ignore environmental context (Zebin Xing et al.).
  • Stochastic Hallucinations: The probabilistic sampling process can produce "noisy" paths that look plausible but are physically unviable (Lang Feng et al.).
  • Semantic Misalignment: Applying constraints to intermediate latent states during generation can distort the learned distribution, leading to unrealistic behavior (Jeongyong Yang et al.).

Integrating Constraints into the Denoising Process

To bridge this gap, researchers are moving beyond "black-box" generation by embedding formal constraints directly into the generative process. This is primarily achieved through three strategies: training-time regularizers, sampling-time guidance, and projected optimization.

Training with Constraint Awareness

One approach is to teach the model about safety during the training phase. Anjian Li et al. introduced a hybrid loss function that penalizes constraint violations while the model is learning. This encourages the model to generate trajectories that serve as high-quality "warm starts" for traditional optimizers, reducing the need for heavy post-processing.

Sampling-Time Guidance and Barrier Functions

A more flexible approach involves guiding a pre-trained model during the inference (denoising) steps. Hao Ma et al. developed CoDiG, which integrates Control Barrier Functions (CBFs) directly into the denoising process. By treating safety as a steering force, the model can adjust its generated path in real-time to avoid obstacles. Similarly, Jeongyong Yang et al. proposed a "prediction-correction" framework that uses CBFs to certify the safety of paths generated via flow matching, achieving significant speedups over iterative diffusion methods.

Denoising as Constrained Optimization

For tasks requiring strict adherence to hard constraints—such as multi-robot coordination—researchers have reframed the denoising step as a constrained optimization problem. Jinhao Liang et al. introduced Simultaneous Multi-Robot Motion Planning (SMD), which uses the Augmented Lagrangian Method to enforce collision avoidance between multiple agents during sampling. Jichen Zhang et al. demonstrated that this optimization-based approach can achieve near-perfect safety in complex robotics tasks without requiring the underlying model to be retrained.

Filtering and Aggregating for Robustness

When real-time optimization is too costly, researchers use filtering and aggregation to manage stochastic risks. Daniele Reda et al. proposed Learned Viability Filters (VFs)—fast Q-functions that evaluate a batch of candidate plans and select the one most likely to succeed. This decouples the generation of natural motion from the evaluation of its safety, allowing the system to discard plans that might lead to delayed failures.

Alternatively, Lang Feng et al. introduced the Trajectory Aggregation Tree (TAT). Instead of relying on a single sampled path, TAT merges multiple samples into a dynamic tree structure. This process filters out the "hallucinated" noise inherent in diffusion models, making the final plan more reliable and allowing the model to work with fewer, faster denoising steps.

Toward Controllable and Safe Simulation

The need for safety-constrained generation extends beyond direct robot control to the creation of training environments. Jingtao Ding et al. developed LCSim, a traffic simulator that uses diffusion models with customizable guide functions to generate realistic but controllable traffic scenarios. By integrating hard safety boundaries into the synthesis mechanism, these tools allow researchers to test autonomous systems against adversarial "edge cases" that are statistically rare but physically plausible, ensuring that the next generation of planners is prepared for the unpredictability of the real world.

Key Foundational Work

CleanDiffuser: An Easy-to-use Modularized Library for Diffusion Models in Decision Making (2024) — contributed 2 concept elements to this direction.
Safe Offline Reinforcement Learning using Trajectory-Level Diffusion Models (2024) — contributed 1 concept elements to this direction.
TAE: A Semi-supervised Controllable Behavior-aware Trajectory Generator and Predictor (2022) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_generative_planning_safety_45809521,
  title = {generative planning safety},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/generative-planning-safety/rdr_458095217b00443d968ce7acd7270ae8},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_generative_planning_safety_45809521_v1,
  title = {generative planning safety},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/generative-planning-safety/rdr_458095217b00443d968ce7acd7270ae8/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake