Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Safety-Critical Control with Offline-Online Neural Network Inference

Junhui Zhang, Sze Zheng Yong, Dimitra Panagou

TL;DR

This work addresses safe motion control for an ego agent operating among other agents with unknown dynamics. It jointly learns the other agents' dynamics offline with radial basis function neural networks and refines the model online using concurrent learning, eliminating the need for persistent excitation. Adaptive conformal prediction provides online, high-probability prediction sets for the learned dynamics, which are embedded into a sampled-data control barrier function framework to guarantee safety with high average confidence. The approach reduces conservatism compared to fixed bounds and demonstrates effective safety guarantees in a multi-agent simulation. The combination of offline-online learning, ACP uncertainty quantification, and CBF-based safety offers a practical path for real-time, safety-critical autonomy in dynamic environments.

Abstract

This paper presents a safety-critical control framework for an ego agent moving among other agents. The approach infers the dynamics of the other agents, and incorporates the inferred quantities into the design of control barrier function (CBF)-based controllers for the ego agent. The inference method combines offline and online learning with radial basis function neural networks (RBFNNs). The RBFNNs are initially trained offline using collected datasets. To enhance the generalization of the RBFNNs, the weights are then updated online with new observations, without requiring persistent excitation conditions in order to enhance the applicability of the method. Additionally, we employ adaptive conformal prediction to quantify the estimation error of the RBFNNs for the other agents' dynamics, generating prediction sets to cover the true value with high probability. Finally, we formulate a CBF-based controller for the ego agent to guarantee safety with the desired confidence level by accounting for the prediction sets of other agents' dynamics in the sampled-data CBF conditions. Simulation results are provided to demonstrate the effectiveness of the proposed method.

Safety-Critical Control with Offline-Online Neural Network Inference

TL;DR

This work addresses safe motion control for an ego agent operating among other agents with unknown dynamics. It jointly learns the other agents' dynamics offline with radial basis function neural networks and refines the model online using concurrent learning, eliminating the need for persistent excitation. Adaptive conformal prediction provides online, high-probability prediction sets for the learned dynamics, which are embedded into a sampled-data control barrier function framework to guarantee safety with high average confidence. The approach reduces conservatism compared to fixed bounds and demonstrates effective safety guarantees in a multi-agent simulation. The combination of offline-online learning, ACP uncertainty quantification, and CBF-based safety offers a practical path for real-time, safety-critical autonomy in dynamic environments.

Abstract

This paper presents a safety-critical control framework for an ego agent moving among other agents. The approach infers the dynamics of the other agents, and incorporates the inferred quantities into the design of control barrier function (CBF)-based controllers for the ego agent. The inference method combines offline and online learning with radial basis function neural networks (RBFNNs). The RBFNNs are initially trained offline using collected datasets. To enhance the generalization of the RBFNNs, the weights are then updated online with new observations, without requiring persistent excitation conditions in order to enhance the applicability of the method. Additionally, we employ adaptive conformal prediction to quantify the estimation error of the RBFNNs for the other agents' dynamics, generating prediction sets to cover the true value with high probability. Finally, we formulate a CBF-based controller for the ego agent to guarantee safety with the desired confidence level by accounting for the prediction sets of other agents' dynamics in the sampled-data CBF conditions. Simulation results are provided to demonstrate the effectiveness of the proposed method.
Paper Structure (13 sections, 2 theorems, 18 equations, 10 figures)

This paper contains 13 sections, 2 theorems, 18 equations, 10 figures.

Table of Contents

  1. Introduction
  2. Preliminaries and Problem Formulation
  3. Preliminaries
  4. Problem Formulation
  5. Methodology
  6. Offline-Online Inference
  7. Offline Learning
  8. Online Learning
  9. Adaptive Conformal Prediction
  10. CBF-based Controller with Inference
  11. Case Study
  12. Conclusions
  13. Acknowledgements

Key Result

Proposition 1

gibbs2021adaptivedixit2023adaptive Consider a time horizon $[t_0,T]$, with sampling time instants $\{t_0, t_1, t_2, \cdots, t_K\}$. For each sampling time $t_k$, where $k\in\{1,2,\cdots,K\}$, the adaptive failure probability $\alpha_i(t_k)$ is updated using recursion Update law of alpha_t with a lea where $\mathcal{P}_i(K)=\frac{\alpha_i(t_0)+\gamma}{K\gamma}$, and $\lim_{K\to +\infty}\mathcal{P}_

Figures (10)

  • Figure 1: The proposed framework of sampled data CBF-based controller with offline-online neural network inference: RBFNNs are utilized to learn other agents' dynamics and are trained in an offline-online manner. Adaptive conformal prediction (ACP) quantifies the estimation error online, and the prediction sets based on ACP are incorporated into the sampled-data CBF-based controller.
  • Figure 2: Estimation of the dynamics of agent 1 in the $x$ direction.
  • Figure 3: Estimation of the dynamics of agent 1 in the $y$ direction.
  • Figure 4: Estimation of the dynamics of agent 2 in the $x$ direction.
  • Figure 5: Estimation of the dynamics of agent 2 in the $y$ direction.
  • ...and 5 more figures

Theorems & Definitions (6)

  • Definition 1
  • Proposition 1
  • Remark 1
  • Remark 2
  • Theorem 1
  • proof