End-to-End Learning of Safe Optimal Feedback Control in High Dimensions with Control Barrier Function Layers

2026-07-22 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

A new methodology addresses the challenge of learning high-dimensional semi-global feedback controllers under hard safety constraints enforced by Control Barrier Functions (CBFs). Prior approaches to embedding quadratic-program-based safety filters as optimization layers in end-to-end policy training were limited to low-dimensional systems, typically with at most 16 state dimensions, due to computational and differentiation bottlenecks. This work overcomes these limitations by combining operator splitting with the Jacobian-Free Backpropagation (JFB) method. This approach enables scalable end-to-end training while preserving hard safety guarantees through the CBF safety filter. The training methodology is theoretically justified using nonsmooth analysis techniques and demonstrated on multi-agent nonlinear control problems, successfully handling systems with state and control dimensions up to 1200 and 400, respectively.

Key takeaway

For Machine Learning Engineers developing safe control systems, this research offers a path to scale end-to-end training for high-dimensional problems. You can now integrate Control Barrier Functions (CBFs) with Jacobian-Free Backpropagation (JFB) to ensure hard safety guarantees without being limited to low-dimensional systems. Consider applying this methodology to multi-agent nonlinear control, potentially handling systems with hundreds of state and control dimensions. This approach enables robust, safe policy learning in complex environments.

Key insights

Scalable end-to-end learning of safe high-dimensional controllers is achieved by integrating CBFs with JFB and operator splitting.

Principles

• Hard safety guarantees are preserved via CBF safety filters.
• Operator splitting and JFB overcome differentiation bottlenecks.
• Nonsmooth analysis justifies the training methodology.

Method

The method combines operator splitting with Jacobian-Free Backpropagation (JFB) to embed a quadratic-program-based CBF safety filter as a differentiable optimization layer for end-to-end policy training.

In practice

• Apply to multi-agent nonlinear control problems.
• Design safe controllers for high-dimensional systems.

Topics

• Control Barrier Functions
• Jacobian-Free Backpropagation
• End-to-End Learning
• High-Dimensional Control
• Multi-Agent Systems
• Optimal Feedback Control

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Robotics Engineer

Related on AIssential

• Safe Reinforcement Learning using Ideas from Model Predictive Control — Combining offline Model Predictive Control with a real-time projection filter enables safe Reinforcement Learning on physical systems.
• BIFROST: Bridging Invariant Feature Representation for Observation-space Sim2Real Transfer — BIFROST enables zero-shot sim2real transfer by learning domain-invariant latent states via a cross-domain bisimulation objective.
• Safety-Guided Flow (SGF): A Unified Framework for Negative Guidance in Safe Generation — Early, strong negative guidance within a critical time window is crucial for safe and high-quality generative model outputs.
• Safe Control using Learned Safety Filters and Adaptive Conformal Inference — ACoFi combines learned safety filters with adaptive conformal inference for robust, scalable control system safety.
• DiRecT: Safe Diffusion-Based Planning via Receding-Horizon Denoising — DiRecT improves diffusion model safety by enforcing constraints only on final trajectories, avoiding intermediate denoising over-constrainment.
• Boosted Control Functions: Distribution Generalization and Invariance in Confounded Models — Boosted Control Functions (BCF) offer strong invariance for prediction under distributional shifts with hidden confounding.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.