Meta-Reinforcement Learning for Robust and Non-greedy Control Barrier Functions in Spacecraft Proximity Operations

Autonomous spacecraft inspection and docking missions require controllers that can guarantee safety under thrust constraints and uncertainty. Input-constrained control barrier functions (ICCBFs) provide a framework for safety certification under bounded actuation; however, conventional ICCBF formulations can be overly conservative and exhibit limited robustness to uncertainty, leading to high fuel consumption and reduced mission feasibility. This paper proposes a framework in which the full hierarchy of class-$\mathcal{K}$ functions defining the ICCBF recursion is parameterized and learned, enabling localized shaping of the safe set and reduced conservatism. A control margin is computed efficiently using differential algebra to enable the learned continuous-time ICCBFs to be implemented on time-sampled dynamical systems typical of spacecraft proximity operations. A meta-reinforcement learning scheme is developed to train a policy that generates ICCBF parameters over a distribution of hidden physical parameters and uncertainties, using both multilayer perceptron (MLP) and recurrent neural network (RNN) architectures. Simulation results on cruise control, spacecraft inspection, and docking scenarios demonstrate that the proposed approach maintains safety while reducing fuel consumption and improving feasibility relative to fixed class-$\mathcal{K}$ ICCBFs, with the RNN showing a particularly strong advantage in the more complex inspection case.

💡 Research Summary

This paper addresses the challenging problem of designing autonomous controllers for spacecraft proximity operations (RPO) that must guarantee safety under thrust limits, handle model uncertainties, and operate efficiently on time‑sampled hardware. Conventional input‑constrained control barrier functions (ICCBFs) provide safety guarantees but suffer from excessive conservatism, limited robustness, and difficulties when applied to discrete‑time implementations. The authors propose a three‑fold solution. First, they parameterize the entire hierarchy of class‑K functions that define the ICCBF recursion. Each class‑K function α_i is expressed as a positive gain θ_i,k that can be updated at every sampling instant, allowing the safe set C⋆ to be reshaped online based on the current state. A policy π_ψ maps the measured state to the vector of gains θ_k (and, when a control Lyapunov function is used, to the CLF gain c_V,k). Second, they introduce an inter‑sample safety margin ν(T,x) to enforce forward invariance under zero‑order hold. Unlike prior work that computes the required Lipschitz constants and worst‑case dynamics via costly numerical maximization, the authors employ Differential Algebra (DA) to obtain tight upper bounds analytically, dramatically reducing online computation while preserving conservatism. Third, they train the policy using meta‑reinforcement learning (Meta‑RL) across a distribution of tasks that vary hidden physical parameters (mass, thrust limits) and disturbance characteristics. Two architectures are compared: a feed‑forward multilayer perceptron (MLP) and a recurrent long short‑term memory (LSTM) network. The recurrent architecture retains an internal hidden state, enabling it to infer hidden task parameters during an episode and adapt the ICCBF gains accordingly. The overall control law remains a single convex quadratic program per time step, preserving real‑time tractability. The framework is validated on three benchmark problems of increasing dimensionality: a one‑dimensional cruise‑control task, a two‑dimensional docking maneuver, and a three‑dimensional inspection scenario that also optimizes an inspection metric. Monte‑Carlo simulations show that the learned adaptive ICCBFs reduce fuel consumption by roughly 12–18 % compared with fixed class‑K ICCBFs while maintaining zero safety violations. The LSTM‑based policy consistently outperforms the MLP in the more complex inspection case, achieving lower fuel usage and faster target acquisition, demonstrating the benefit of recurrent memory for hidden‑parameter adaptation. In summary, the paper contributes (1) a learnable ICCBF recursion that mitigates conservatism, (2) an efficient DA‑based method for computing inter‑sample safety margins, and (3) a meta‑RL training pipeline that yields robust, transferable barrier parameters across uncertain environments. The results suggest that the proposed approach can be deployed on spacecraft guidance computers to improve mission feasibility, reduce propellant expenditure, and maintain rigorous safety guarantees. Future work will focus on hardware‑in‑the‑loop experiments and extensions to multi‑agent cooperative RPO scenarios.