💻 computer science

Motion Planning with Model-Based Diffusion via Constraint Optimization and Adaptive Scheduling

This paper introduces MD-COAS, a novel single-robot motion planning framework that unifies soft and hard constraint enforcement via an inexact Augmented Lagrangian Method and Convex Feasible Set projection within an adaptive diffusion scheduling scheme, achieving superior safety, success rates, and convergence on complex non-convex benchmarks compared to existing model-based diffusion approaches.

Original authors: Zhilin He, Bowei Li, Jianlin Dou, Yuner Zhang, Changliu Liu

Published 2026-07-17

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Zhilin He, Bowei Li, Jianlin Dou, Yuner Zhang, Changliu Liu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot how to walk through a crowded, chaotic room filled with furniture, moving people, and narrow doorways. The goal is simple: get from point A to point B without bumping into anything. But the math behind this is incredibly hard. The robot has to figure out a path that doesn't just avoid walls, but also obeys the laws of physics (it can't teleport or turn on a dime) and finishes the job quickly. This field is called Motion Planning.

For a long time, robots used to map out every possible path like a giant spiderweb before choosing one, or they used to learn by watching thousands of videos of humans doing the task. But there's a new, clever way to do this called Diffusion. Think of diffusion like a game of "Hot and Cold" or a messy sketch that gets cleaned up. You start with a completely random, jumbled scribble (pure noise) and slowly, step-by-step, refine it into a smooth, perfect line. The robot uses this "cleaning" process to turn a chaotic guess into a safe, working path. The big challenge, however, is making sure that as the robot "cleans up" its path, it doesn't accidentally erase the safety rules, like the rule that says "do not hit the table."

This paper introduces a new method called MD-COAS (Model-Based Diffusion via Constraint Optimization and Adaptive Scheduling) that acts like a super-smart editor for these robot paths. The authors found that previous methods were a bit too rigid or too loose: some tried to gently nudge the robot away from walls (soft rules), while others forced the robot to snap back to safety instantly (hard rules), but neither worked perfectly when the room was extremely cluttered. MD-COAS solves this by combining the best of both worlds. It uses a "soft" mathematical guide to gently steer the robot toward good paths while keeping a "hard" safety net ready to catch it if it gets too close to a crash.

But the real magic is in the Adaptive Scheduling. Imagine you are editing a rough draft of a story. You wouldn't spend the same amount of time checking every single sentence. You might skim the easy parts but spend hours polishing the tricky scenes. Similarly, MD-COAS doesn't treat every step of the robot's path-finding the same way. It watches how "messy" the path is at each moment. If the robot is doing fine, the system relaxes and saves energy. But if the robot starts drifting toward a wall, the system instantly tightens its grip, applying more intense safety checks and using more computing power exactly when and where it's needed.

In their tests, the researchers simulated this robot in two different worlds. First, they used a simple 2D map with obstacles that looked like a chaotic maze of circles and squares, some of which were merged into weird, non-convex shapes (like a C-shape or a U-shape). They also tested it on a complex 7-DoF (seven degrees of freedom) robotic arm, which is like a human arm with a shoulder, elbow, and wrist that can twist and turn in many directions.

The results showed that MD-COAS was the clear winner. In the messy 2D mazes, while other methods often got stuck or crashed into obstacles, MD-COAS successfully found safe paths almost every time, even in the tightest spots. It also found paths that were cheaper (shorter or faster) and converged to the solution faster than the competition. When they moved to the 7-DoF robotic arm, the method continued to shine. Even as the obstacles got bigger and the safe space shrank, MD-COAS maintained a 100% success rate with zero collisions, whereas other methods started to fail or miss the target.

The paper suggests that this success comes from not treating safety as a static rulebook. Instead, by dynamically adjusting how strictly it enforces safety and how much computing power it spends on corrections, the robot can explore more creative paths without ever losing its way. It's a bit like having a coach who knows exactly when to let the athlete run free and when to step in and correct their form, ensuring they finish the race strong and safe.

Technical Summary: Model-Based Diffusion via Constraint Optimization and Adaptive Scheduling (MD-COAS)

Problem Statement

Single-Robot Motion Planning (SRMP) in highly non-convex constrained environments presents a significant challenge. Robots must generate trajectories that satisfy collision-free guarantees, dynamic feasibility, and task-related constraints while operating under computational limits.

Existing approaches face specific limitations:

Data-Driven Diffusion: Methods relying on demonstration data (e.g., MPD) bypass global geometric search but require large-scale datasets and often fail to rigorously enforce dynamics within the generative model.
Model-Based Diffusion (MBD): While MBD incorporates explicit dynamics to estimate scores analytically without demonstrations, existing safety-enhanced variants struggle in highly non-convex environments. They typically rely on either soft feasibility priors (which may stall near tight passages) or hard projection operators (which may be computationally expensive if applied uniformly). Furthermore, these methods often fix safety enforcement parameters throughout the diffusion process, neglecting the changing nature of the denoising schedule.

Methodology

The paper proposes Model-Based Diffusion via Constraint Optimization and Adaptive Scheduling (MD-COAS). This framework formulates SRMP as a trajectory optimization problem where the reverse diffusion process acts as a solver. The method integrates three core components:

1. Soft Feasibility via Inexact Augmented Lagrangian Method (iALM)

To avoid the statistical inefficiency of hard indicator functions (which create sparse weighting landscapes), the authors replace the hard feasibility prior with a soft kernel based on the Inexact Augmented Lagrangian Method (iALM).

The target distribution is defined as $p(\tau) \propto \exp(-\tilde{J}(\tau)/\beta)$ , where $\tilde{J}$ includes the task objective plus a penalty term for constraint violations.
The penalty term uses dual variables ( $\lambda$ ) and a penalty parameter ( $\rho$ ) to assign continuous, non-zero weights to mildly infeasible trajectories, preventing "dead samples" while guiding the diffusion toward feasible regions.

2. Hard Projection via Convex Feasible Set (CFS)

To ensure strict safety in highly non-convex environments where soft guidance may stall, MD-COAS employs a Convex Feasible Set (CFS) based hard projection operator ( $P_{I,H}$ ).

This operator projects nominal diffusion proposals onto a local convex inner approximation of the safe set.
It iteratively convexifies collision-avoidance constraints and solves a structured Quadratic Program (QP) over the full control sequence to find a minimal-perturbation correction.
This acts as a proximal correction step, enforcing local safety guarantees that soft priors alone cannot provide.

3. Adaptive Scheduling (AS)

A key innovation is the adaptive scheduling of both the soft prior parameters and the hard projection operator, co-optimized within a primal-dual view.

Soft Prior Adaptation: The iALM parameters ( $\lambda_k, \rho_k$ ) are updated at each reverse diffusion step based on a robust batch statistic of constraint violations ( $\tilde{r}_k$ ). If violations persist, feasibility pressure increases; if they subside, parameters relax.
Hard Operator Adaptation: Instead of applying the computationally expensive CFS projection at every step, the method uses a gating mechanism. The probability of projection ( $p_k$ ), the active set size ( $H_k$ ), and the iteration budget ( $I_k$ ) are dynamically adjusted based on a compute dual variable ( $\nu_k$ ).
Budget Constraint: The system enforces a fixed computational budget ( $B$ ) per diffusion step. If the realized effort exceeds the budget, the dual variable $\nu_k$ increases to discourage aggressive projection; if the budget is underutilized, $\nu_k$ decreases to allow more frequent corrections.

Key Contributions

The paper outlines two primary contributions:

Unified Constraint Optimization Framework: The development of a framework that integrates the iALM soft diffusion prior with a CFS-based hard projection operator. This combination preserves informative diffusion guidance and multimodality while certifying safety in highly non-convex environments.
Adaptive Scheduling Strategy: The proposal of a primal-dual adaptive scheduling mechanism (MD-COAS) that co-optimizes safety enforcement parameters and the diffusion schedule. This allows the system to adaptively regulate constraint enforcement based on the current state of the diffusion annealing process and available computational resources.

Experimental Results

The authors evaluated MD-COAS on two benchmarks: randomly generated 2D constrained obstacle environments and a 7-DoF robotic arm avoidance task (D3IL).

2D Benchmarks: MD-COAS achieved higher Safe & Success Rates (SSR) and lower final costs compared to baselines (MBD, EB-MBD, MDOC, SafeDiffuser, DPCC-C).
- In highly non-convex "Union Maps," MD-COAS maintained an SSR of ~0.5, outperforming other methods which degraded significantly (e.g., EB-MBD dropped to ~0.7, MDOC to ~0.5).
- Ablation studies showed that while soft priors alone (MD-COAS-A) converged, they suffered from higher terminal costs and reduced SSR in tight environments. Conversely, fixed hard projection (MD-COAS-F) achieved high safety but at a higher computational cost. MD-COAS balanced these by achieving comparable safety to the fixed version with improved runtime efficiency.
7-DoF Arm Task: In the higher-dimensional joint-space control task with varying obstacle radii, MD-COAS consistently achieved 100% success with zero violations across all difficulty levels.
- Other methods showed reduced robustness as obstacle size increased.
- MD-COAS preserved trajectory diversity (multimodality) better than MDOC, which tended to collapse into a single mode due to overly restrictive feasible sets.

Significance and Claims

The paper claims that MD-COAS addresses the critical gap in existing diffusion-based planners: the lack of a unified framework that combines soft and hard constraint enforcement with adaptive resource management.

Robustness: By unifying iALM and CFS, the method handles the "stalling" of soft priors in tight passages while avoiding the computational inefficiency of uniform hard projections.
Efficiency: The adaptive scheduling allows the planner to concentrate computational effort on diffusion steps where constraint violations are most severe, improving runtime without sacrificing solution quality.
Generality: The method demonstrates effectiveness across varying dimensions (2D to 7-DoF) and constraint complexities (convex to highly non-convex), offering a principled approach to safe, model-based motion planning without reliance on large-scale demonstration datasets.

The authors conclude that future work will focus on improving sampling projection efficiency, adapting the planner to tasks with richer dynamics, and exploring multi-robot coordination settings.