Safety Certificate against Latent Variables with Partially Unidentifiable Dynamics
Authors: Haoming Jing, Yorie Nakahira
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The effectiveness of the proposed safety certificate is demonstrated in numerical simulations. |
| Researcher Affiliation | Academia | 1Electrical and Computer Engineering Department, Carnegie Mellon University, Pittsburgh, USA. Correspondence to: Yorie Nakahira <EMAIL>. |
| Pseudocode | Yes | Algorithm 1 Generation of D using D Algorithm 2 Proposed control algorithm |
| Open Source Code | No | The paper does not contain any explicit statements or links indicating that source code for the described methodology is publicly available. |
| Open Datasets | No | Let D = {D1, D2, , DND} denote the available offline data. Here, ND is the size of the training dataset, and each individual data Di, i {1, 2, , ND} contains the sequence of observable state {Xt}t {0,1, ,H} := X0:H and control action U0:H in an episode. The offline dataset can be considered as human driving dataset where the human observes the slipperiness of the road in their behavioral policy, but the slipperiness is not recorded by the sensor. |
| Dataset Splits | No | The paper describes using an 'offline dataset D' and creating a 'dataset D' from it for learning. It mentions running 100 simulations with 100 trajectories for evaluation, but does not specify any training/validation/test splits for machine learning models or how these trajectories were partitioned for experimental evaluation. |
| Hardware Specification | No | The paper mentions 'numerical simulations' but does not provide any specific details about the hardware (e.g., CPU, GPU models, memory) used to run these simulations or experiments. |
| Software Dependencies | No | The paper references various causal reinforcement learning methods and control barrier function techniques by citing other papers (e.g., Wang et al. 2021b, Shi et al. 2024, Cosner et al. 2023) but does not list specific software libraries, frameworks, or their version numbers that were used for implementation. |
| Experiment Setup | Yes | We consider a setting that resembles a simplified driving scenario with discrete state space. Let Xt = [X1 t , X2 t ]T Z2 be the state of the system, where X1 t represents the position of the vehicle on a 1-dimensional road, and X2 t represents the velocity of the vehicle. The control action Ut { 3, 2, 1, 0, 1} represents the acceleration or deceleration applied to the wheels. The latent variable Wt {0, 1, 2, 3} represents the slipperiness of the road, which can make the acceleration or deceleration applied to the wheels less effective. The system also has uncertainty Nt = [N 1 t , N 2 t ]T { 1, 0, 1} { 2, 1, 0, 1, 2}. The system transition is given by X1 t+1 =X1 t + X2 t (56) X2 t+1 = max(0, X2 t + sign(Ut + N 1 t ) max(0, |Ut + N 1 t | Wt) + N 2 t ). (57) ... We consider H = 10 and ϵ = 0.2. We run 100 simulations, where each simulation simulates 100 trajectories starting from X0 = [0, 0]T... For the discrete-time control barrier function, we first represent the safety requirement using C(Xt) = 1{Xt C}, where C = {x Z2 : h(x) 0}, and h([x1, x2]T ) = tanh n {1,3,5,7} 4 nπ sin( π 5 n(x1 + 0.5)) x2 We use the safety condition E[h(Xt+1)|Xt, Ut] αh(Xt) + δ (119) with α = 0.01 and δ = 0.5, such that the condition (6) is guaranteed for ϵ = 0.2 due to Cosner et al. 2023, equation (13). |