Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Safe and Stable Control via Lyapunov-Guided Diffusion Models
Authors: Xiaoyuan Cheng, Xiaohang Tang, Yiming Yang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To validate our approach, we conduct experiments on a wide variety of dynamical control systems, where S2Diff consistently outperforms both certificatebased controllers and model-based diffusion baselines in terms of safety, stability, and overall control performance. |
| Researcher Affiliation | Academia | Xiaoyuan Cheng Xiaohang Tang Yiming Yang University College London, United Kingdom |
| Pseudocode | Yes | Algorithm 1: S2Diff Input: Distribution of initial states Dx0, model dynamics f, nominal policy, number of training epochs K Output: Certificate function (CLBF) VK for epoch k = 1 to K do // === Phase 1: Guided Trajectory Sampling === Initialize an empty dataset of new trajectories D; for each initial state x0 in a batch sampled from Dx0 do Generate one full trajectory via a guided denoising process by maximizing Equations (6) and (8); Sample a clean trajectory U 0 by applying the reverse diffusion process (Equation (10)) with model dynamics f, starting from noise; The process is guided at each step by the current CLBF Vk 1; Add the resulting trajectory U 0 to the dataset D; // === Phase 2: CLBF Update === Use the entire newly generated dataset D for training; Update CLBF parameters by performing gradient descent on the loss from Equation (11), using trajectories from D, obtain Vk; |
| Open Source Code | Yes | 5. Open access to data and code Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes] Justification: See supplementary materials. |
| Open Datasets | No | In this section, we comprehensively evaluate the performance of our model across various nonlinear dynamical systems, covering both tracking and control tasks. ... (1) Inverted pendulum (n = 2): stabilize the pendulum in the upright position by moving the base left or right. ... (6) F-16 (n = 16): control a F-16 aircraft to stabilize key variables such as altitude, velocity, angle of attack, elevator position and other critical states. The paper describes the *models* and *environments*, not pre-collected *datasets* with access links. |
| Dataset Splits | No | We evaluate the control policies from three aspects: (1) Safety rate the percentage of trajectories that remain safe, computed over 20 initial states; The experiments are based on simulated environments and different initial states, not a pre-defined dataset with splits. |
| Hardware Specification | Yes | Note that all inference time evaluations were conducted on the same device: Intel i9-13900 CPU with one RTX 4090 GPU. |
| Software Dependencies | No | MPC. This baseline employs model predictive control (MPC), solved using the Gurobi optimizer at each time step. The text mentions "Gurobi optimizer" but does not provide a specific version number, which is required by the strict rule for this question. |
| Experiment Setup | Yes | Table 7: Hyperparameter configurations for eight tasks using S2Diff. Task NN Architecture Safe Level c Temp. Loss Coeff. Traj. Len. Inverted Pendulum 3 64 1 0.1 1,1 5 Car (Kin.) 3 64 1 0.1 1,1 5 Car (Slip) 3 64 1 0.1 1,1 5 Segway 3 64 1 0.1 1,1 5 Neural Lander 3 64 10 0.1 1,1 5 2D Quad 3 64 1 0.1 1,1 5 3D Quad 3 64 10 0.1 1,1 5 F-16 3 128 10 0.1 1,1 5 |