ABNet: Attention BarrierNet for Safe and Scalable Robot Learning

TL;DR

The paper addresses the scalability and training stability challenges of barrier-based safe robot learning by introducing Attention BarrierNet (ABNet), a multi-head framework where each head learns a safety-critical policy using distinct observation features. Safety is preserved by defining new High-Order Control Barrier Function (HOCBF) constraints over the fused head outputs $\bm u = \sum_{k=1}^h w_k \bm u_k$, with trainable weights $w_k$ that sum to unity, and by solving differentiable quadratic programs in each head. The authors provide formal safety guarantees (via Nagumo's theorem) and demonstrate through 2D obstacle avoidance, safe manipulation, and vision-based driving that ABNet achieves superior robustness and lower control uncertainty compared with baselines, including single BarrierNets and end-to-end models. The approach enables incremental scaling of safe models, maintains safety during merging, and shows practical impact for safe autonomous robotic systems in noisy environments.

Abstract

Safe learning is central to AI-enabled robots where a single failure may lead to catastrophic results. Barrier-based method is one of the dominant approaches for safe robot learning. However, this method is not scalable, hard to train, and tends to generate unstable signals under noisy inputs that are challenging to be deployed for robots. To address these challenges, we propose a novel Attention BarrierNet (ABNet) that is scalable to build larger foundational safe models in an incremental manner. Each head of BarrierNet in the ABNet could learn safe robot control policies from different features and focus on specific part of the observation. In this way, we do not need to one-shotly construct a large model for complex tasks, which significantly facilitates the training of the model while ensuring its stable output. Most importantly, we can still formally prove the safety guarantees of the ABNet. We demonstrate the strength of ABNet in 2D robot obstacle avoidance, safe robot manipulation, and vision-based end-to-end autonomous driving, with results showing much better robustness and guarantees over existing models.

Figures (9)

• Figure 1: The proposed ABNet that is robust, scalable and generates stable output while guaranteeing safety for robots. Each head of BarrierNet in the model could learn safe control policies with attention on different observation feature in a scalable or one-shot manner.
• Figure 2: Architecture of multi-head BarrierNets (i.e., ABNet). The ABNet is usually used in conjunction with any other neural networks and can be implemented in parallel. The parameters (inputs) of each head of BarrierNet are the outputs of previous layers (such as CNN or LSTM).
• Figure 3: 2D robot obstacle avoidance closed-loop testing control profiles (left) and ABNet performance with the increasing of BarrierNet heads using scalable training (right). This scalable training for ABNet is with safety guarantees. The controls are subject to input noise, and thus are non-smooth.
• Figure 4: Robot manipulation closed-loop end-effector trajectories (left) and ABNet performance with the increasing of BarrierNet heads using scalable training (right). This scalable training for ABNet is with safety guarantees. The transparent red and blue trajectories in the left figure are corresponding to BNet and ABNet-100 models in all runs, respectively.
• Figure 5: Vision-based end-to-end autonomous driving closed-loop testing trajectories in VISTA (left) and ABNet performance with the increasing of BarrierNet heads using scalable training (right). This scalable training is done by both the ABNet and ABNet-att in Table \ref{['tab:comp_driving']} with safety guarantees.

Theorems & Definitions (7)

• Definition 2.1
• Definition 2.2
• Definition 2.3
• Definition 2.4
• Theorem 2.5: Xiao2021TAC2
• Theorem 3.1
• Theorem 3.2