Non-Myopic Safety Filtering with Learned Intervention Cost-to-Go

Johannes Autenrieb, Mark Spiller, Daniel Silvestre, Namhoon Cho, Jason J. ChoiarXiv 2026 · arXiv

TL;DR

Standard safety filters are "myopic," minimizing only the immediate correction needed to stay safe, which often leads to massive corrections later. This paper introduces a non-myopic filter that uses a learned "cost-to-go" function to anticipate future interventions, reducing cumulative control effort by 50% while maintaining rigorous safety guarantees.

What Problem Does This Paper Address?

In autonomous systems like drones or robots, Control Barrier Functions (CBFs) are the gold standard for safety. They act as a "safety filter," minimally adjusting a user's command to ensure the system stays within a safe set (e.g., not hitting a wall).

However, traditional CBF filters are myopic. They only care about the instantaneous deviation from the nominal command. Imagine driving toward a sharp turn: a myopic filter might let you maintain high speed until the very last millisecond because you are technically "safe" right now. This forces a violent, high-energy intervention at the last moment.

The authors tackle this "greedy" behavior. They want a filter that can look ahead and say, "If I make a small, proactive correction now, I can avoid a huge, costly intervention later."

Methods

The authors propose the Non-Myopic Safety Filter (NMF). The core innovation is incorporating a learned approximation of the optimal intervention cost-to-go.

  1. Learned Value Function ($W_\theta$): They train a neural network to estimate the total future intervention cost from any given state. This function represents the "price" of safety over the long term.
  2. Relaxed Bellman-Residual Constraint: Instead of just checking the barrier condition, the filter includes a constraint based on the Bellman equation. It tries to ensure that the current intervention plus the estimated future cost is consistent with the current state's estimated cost.
  3. Hard vs. Soft Constraints: Crucially, the safety constraint (CBF) remains a hard constraint. The learned cost-to-go is treated as a relaxed constraint with a "slack" variable ($\delta$). This ensures that even if the neural network makes a poor prediction, the system remains safe and the optimization problem remains feasible.
  4. Single-Step Optimization: Unlike Model Predictive Control (MPC), which solves a complex multi-step trajectory optimization at every tick, NMF solves a single-step Quadratic Program (or a close approximation). This makes it computationally much lighter.

Datasets

The researchers evaluated their approach in a simulated 3D quadrotor obstacle-avoidance task.

  • Training Data: They generated 7,619 "continuation-cost" labels from 800 training starts. These labels were created by solving an offline optimal control problem to find the true minimum intervention cost.
  • Evaluation: They tested the filter on 100 initial conditions that were entirely unseen during training or parameter tuning.
  • Baselines: They compared NMF against standard CBF, a Hamiltonian-based filter (H-CBF-QP), and a high-performance MPC-CBF with a 15-step lookahead.

Key Insights and Ideas

  • Preventive Intervention: The NMF is willing to intervene even when the current nominal command is perfectly safe. By accepting a small "stage cost" now, it drastically reduces the "cost-to-go."
  • Safety-Performance Decoupling: Because the learned value function enters through a relaxed constraint, the safety of the system does not depend on the accuracy of the neural network. If the network is wrong, the slack variable absorbs the error, but the hard CBF constraint still prevents collisions.
  • The "Aha" Moment: The authors proved a "Pathwise Performance Certificate." This means they can provide a mathematical bound on how much intervention effort will be used over a trajectory using only the learned values and the realized slack, without needing to know the "true" optimal value function.

Key Results

  • 50% Effort Reduction: NMF reduced the mean cumulative intervention cost by approximately 49.5% compared to conventional CBF filtering.
  • Consistency: The performance was stable across three independently trained networks with different random initializations.
  • Speed: NMF was significantly faster than the predictive MPC-CBF baseline, requiring only about 13% of the online computation time per step.
  • Better Safety Margins: Qualitatively, the NMF trajectories maintained larger clearances from obstacles (e.g., increasing minimum clearance from 0.062m to 0.217m) because it "saw" the danger coming earlier.

Limitations and Future Work

  • Model Dependency: The current formulation assumes the system dynamics ($f$ and $g_u$) are known. Future work aims to adapt the filter for environments with model uncertainty.
  • Fixed Environments: The value function was trained for a specific obstacle geometry. Scaling this to dynamic or unknown environments would likely require more generalized features or online learning.
  • Suboptimality Bounds: While the paper provides a suboptimality bound, it includes terms that grow with the time horizon, meaning the gap between NMF and the theoretical optimum isn't yet guaranteed to be horizon-independent.

Summary updated

Abstract

Control barrier function (CBF) safety filters minimize instantaneous deviation from a nominal control command, but a sequence of pointwise-minimal interventions need not minimize cumulative intervention effort. This paper proposes a discrete-time non-myopic safety filter using a learned approximation of the optimal intervention cost-to-go over CBF-admissible policies. The approximation enters through a relaxed Bellman-residual constraint, while the original barrier condition remains hard. Feasibility is equivalent to that of the conventional CBF filter, and forward invariance remains independent of approximation accuracy under the stated assumptions. We derive a trajectory-wise performance certificate that requires neither the optimal value function nor an approximation-error bound. Under additional approximation and horizon assumptions, a suboptimality bound contains an explicit endpoint approximation-error term together with accumulated tolerance and relaxation terms. Numerical results for a collision-avoidance problem show an approximately 50% reduction in mean cumulative intervention effort relative to conventional CBF filtering over 100 initial conditions not used during training or parameter selection. The improvement is consistent across three independently trained networks, while maintaining the prescribed safety constraints and requiring less online computation than the tested predictive filter.
BibTeX
@misc{autenrieb2026non,
  author = {Johannes Autenrieb and Mark Spiller and Daniel Silvestre and Namhoon Cho and Jason J. Choi},
  title = {{Non-Myopic Safety Filtering with Learned Intervention Cost-to-Go}},
  year = {2026},
  eprint = {2611.21036},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.21036},
}
Figure