IEEE/RSJ IROS 2026 · Pittsburgh, PA, USA

D-SafeMPC: Diffusion-Driven Safe Model Predictive Control with Discrete-Time Control Barrier Functions

Erdi Sayar¹ Ersin Daş² Joel W. Burdick³ Alois Knoll⁴ Erdal Kayacan¹

¹ Paderborn University, Germany ² Illinois Institute of Technology, Chicago, USA ³ California Institute of Technology (Caltech), Pasadena, USA ⁴ Technical University of Munich (TUM), Germany

Code Repository (GitHub) Abstract Architecture Robot Experiments Cite (BibTeX)

D-SafeMPC Framework Architecture

Figure 1: Overview of the D-SafeMPC trajectory planning framework. At each denoising step, the reverse diffusion process is guided by discrete-time Control Barrier Function (CBF) and Control Lyapunov Function (CLF) gradients ($\nabla_{\mathbf{u}}\psi_h$ and $\nabla_{\mathbf{u}}\Phi_V$), steering candidate trajectories away from obstacles and toward the goal. The trajectory is then projected onto a safe and dynamically feasible set via an MPC refinement step, which in turn initializes the next denoising iteration.

Abstract & Key Contributions

A key limitation on the use of diffusion models in robotic trajectory planning is their inability to inherently enforce hard safety or dynamical constraints, which often results in physically infeasible or unsafe outputs. Hybrid approaches that employ model predictive control (MPC) to address this problem can be unstable, as poor trajectory initializations from the diffusion model prevent the MPC from converging to a safe and feasible solution.

To overcome these challenges, we propose D-SafeMPC, which fundamentally enhances the interaction between diffusion models and optimal control. Our method guides the reverse diffusion process with control barrier functions (CBFs) and control Lyapunov functions (CLFs), and employs an iterative-projection scheme where an MPC refines the trajectory at each denoising step. This steers sampling toward safe, goal-directed regions and provides reliable MPC warm starts.

In extensive simulations on a 7-DOF Franka manipulator across four scenarios (one static-obstacle and three dynamic-obstacle settings) and in a sim-to-real experiment on a physical Franka robot, D-SafeMPC improves safety, task success rates, and planning efficiency over state-of-the-art baselines.

Discrete-Time CBF/CLF Guidance

Guides the stochastic reverse diffusion process with discrete-time CBF and CLF gradients, ensuring high-quality warm-starts without retraining the diffusion backbone.

Iterative-Projection MPC

Enforces strict hard safety and kinematic/dynamic feasibility constraints by projecting trajectories onto admissible control sets at each denoising step.

Real-World Franka Sim2Real

Validated on a physical 7-DOF Franka Emika Panda robot with real-time avoidance of moving obstacles, demonstrating zero collisions and fast execution.

Physical Franka Robot Experiments

Real-Time Dynamic Obstacle Avoidance

We deployed D-SafeMPC zero-shot on a physical 7-DOF Franka Emika Panda manipulator. The robot successfully executes reaching tasks while evading dynamic obstacle intrusions:

Hardware Setup

7-DOF Franka Emika robot reaching across workspace under hard safety constraints.

Dynamic Evasion

Rapid trajectory re-planning when an unexpected dynamic obstacle crosses the path.

Simulation Benchmarks & Baselines

Static and Dynamic Obstacle Scenarios

We benchmarked D-SafeMPC against Diffuser, CoBL-Diffusion, and DPCC across four demanding scenarios in MuJoCo:

(a) Static Obstacle

Constrained corridors with immovable barriers.

(b) Dynamic Obstacles

Oscillating and moving obstacles with time-varying barrier functions.

Open-Source Code & Configurations

Full PyTorch implementation, MPC projection formulation, Franka robot controllers, and experiment configs:

View Repository (github.com/erdiphd/D-SafeMPC)

BibTeX Citation

If you find D-SafeMPC or this framework useful for your research, please cite:

@inproceedings{sayar2026dsafempc,
  title     = {D-SafeMPC: Diffusion-Driven Safe Model Predictive Control with Discrete-Time Control Barrier Functions},
  author    = {Sayar, Erdi and Da{\c{s}}, Ersin and Burdick, Joel W. and Knoll, Alois and Kayacan, Erdal},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  year      = {2026},
  address   = {Pittsburgh, PA, USA},
  publisher = {IEEE}
}