Safe-Reachability Objectives: Theory & Methods

• Safe-Reachability Objectives are a formal paradigm combining liveness (target reachability) and safety (avoidance of failure regions) in dynamic and stochastic systems.
• They drive applications in control theory, formal verification, reinforcement learning, hybrid systems, and multi-agent games for safety-critical autonomy.
• Methodologies include Hamilton–Jacobi analysis, dynamic programming, SMT-based policy synthesis, and data-driven neural approaches to ensure robust safety and performance.

Safe-reachability objectives formalize the requirement that a system must achieve liveness goals (e.g., reach a target set) while maintaining specified safety constraints (e.g., always avoid failure or unsafe regions). This paradigm appears across control theory, formal verification, reinforcement learning, hybrid systems, stochastic planning, and multi-agent games, serving as a mathematical and computational foundation for safety-critical autonomy.

1. Conceptual Foundations and Formal Definitions

Safe-reachability combines “reachability” (guaranteed arrival at a goal set) with “safety” (invariant avoidance of unsafe sets). In classical dynamical systems, the problem is typically stated as: find a control policy so that for every admissible disturbance and initial condition within a set $S_0$, the state trajectory $x(t)$ satisfies:

• $\exists\,t\leq T:\;x(t)\in G$ (reach the goal $G$ within horizon $T$) and
• $\forall\,t\in[0,T]:\;x(t)\notin F$ (never visit the “failure” or unsafe set $F$)

This reach–avoid property is central in modern safety-critical AI, multi-agent planning, reinforcement learning, hybrid systems, stochastic games, and model checking. The corresponding mathematical sets and value functions are called reach–avoid sets or safe–reachability sets (Hsu et al., 2021, Wang et al., 2018, Chen et al., 2015).

More generally, for discrete-time systems with state space $S$, action space $U$, goal set $G$, and unsafe set $F$, the reach–avoid set is

$$\mathcal{RA}(G, F) = \{\, x_0 \mid \exists\,u_\cdot,\; \exists\,\tau\geq0\;\mathrm{s.t.}\; x_\tau\in G,\; x_k\notin F \;\forall\,k\leq\tau \,\}$$

(Hsu et al., 2021).

In stochastic settings (e.g., MDPs, POMDPs), safe-reachability may be stated as ensuring

• Probability to reach $G$ exceeds a threshold, while
• Probability to visit $F$ never exceeds another threshold

(Ganai et al., 2023, Wang et al., 2018).

2. Mathematical Methods and Theoretical Frameworks

a. Hamilton–Jacobi Reachability (HJR)

Hamilton–Jacobi (HJ) methods cast reach–avoid as a (possibly differential-game) variational inequality over a value function, often of the form:

$$\min\left\{ \frac{\partial V}{\partial t} + H(x, \nabla V),\; \ell(x) - V(x, t) \right\} = 0$$

with appropriate terminal conditions, where $\ell(x)$ encodes the unsafe (zero-level) or goal set (Hsu et al., 2021, Lin et al., 2022, Chen et al., 2015, Chen et al., 2021). The backward reachable tube is identified with sublevel, typically $\{ x : V(x,0) \le 0 \}$.

For liveness (goal-reach), a Hamilton–Jacobi–Bellman PDE is solved; for safety (avoidance), Hamilton–Jacobi–Isaacs equations appear.

b. Dynamic Programming and Bellman Recursions

In discrete or stochastic systems, safe-reachability is encoded by min-max recursions (“Bellman equations”) over value or Q-functions, e.g.,

$$V(x) = (1-\gamma)\max\{\ell(x),g(x)\} + \gamma \max\left\{ g(x), \min\{ \ell(x), \inf_{u} V(f(x, u)) \} \right\}$$

with discount factor $\gamma$, liveness margin $\ell$, and failure margin $g$ (Hsu et al., 2021, Ganai et al., 2023).

Probabilistic safe-reachability in RL introduces Reachability Estimation Functions (REFs), recursively defined as

$$\phi^\pi(s) = \max\left\{ \mathbf{1}_{s\in S_v}, \;\mathbb{E}_{a,s'} [ \phi^\pi(s') ] \right\}$$

capturing the future violation probability under policy $\pi$ (Ganai et al., 2023).

c. Set-valued Analysis: Forward/Backward Reachable Sets

In continuous control, reachable sets (FRS/BRS) describe the collection of all states that the system can (under policy and adversarial disturbance) be driven to (or avoided). This is used both for deterministic and data-driven dynamics (Kousik et al., 2019, Holmes et al., 2020, Hafez et al., 5 Mar 2025).

For hybrid systems, classical “finite-step” reachability can be unsafe (under-approximate) in the presence of Zeno behaviors. To address this, “safe reachability” is defined as the minimal closed set containing the finite-step reach set and all its limits, ensuring over-approximation and robustness (Moggi et al., 2017).

3. Algorithmic Synthesis and Implementation

Numerous algorithms have been developed, including both exact and approximate methods:

a. Receding-Horizon and Real-Time Safe Trajectory Synthesis

In receding-horizon safe planning (e.g., quadrotor, manipulator, or multi-robot), offline-computed reachable sets or over-approximating zonotopes parameterize the set of safe plans. At runtime, only plans whose entire reachable-tube avoid obstacles are eligible; otherwise, fail-safe backups (hovering, braking) are triggered. Example procedures include parameter elimination via zonotope intersection (Kousik et al., 2019, Holmes et al., 2020).

b. SMT-Based Policy Synthesis for POMDPs

Safe-reachability objectives in POMDPs are realized by encoding the goal-constrained belief space as a symbolic constraint system, incrementally search for a policy using Satisfiability Modulo Theories (SMT) solvers (Wang et al., 2018). This reduces the intractable full belief-exploration to manageable optimization over a subspace.

c. Deep Learning and Neural Approximation

Neural PDE solvers, such as DeepReach, are trained to solve high-dimensional HJ reachability VIs. To address possible optimism in the neural solution, scenario-based error-certification uses sampling and statistical bounds to ensure corrected value-functions yield probabilistically safe reachable tubes (Lin et al., 2022, Nakamura et al., 2 Feb 2025). For tasks with raw image input, safety filtering is performed directly in the latent space of a learned world-model with reachability-theoretic backup (Nakamura et al., 2 Feb 2025).

d. Safe RL and Supervisory Control

Safe RL algorithms (e.g., reach-avoid Q-learning, RESPO) optimize reward within the subset of states certifiably free of violations and conservatively minimize cost elsewhere. Where the safety critic is only an approximation, a runtime “shield” invokes backup safety policies on untrusted actions (Hsu et al., 2021, Ganai et al., 2023, Chen et al., 2021).

e. Data-driven Reachability for Black-Box Systems

Without analytical models, data-driven reachability employs local regression, estimated Lipschitz bounds, and set-based (zonotopic) over-approximation to verify that LLM-proposed (or teleoperator) maneuvers are provably safe, with fallback plans adjusted via projected gradient steps (Hafez et al., 5 Mar 2025).

4. Safe-Reachability in Games and Multi-Agent Systems

Safe-reachability objectives have been extensively studied in multi-player games, including turn-based, stochastic, and lexicographically ordered games:

• In quantitative reachability/safety games, existence of finite-memory Nash equilibria and secure equilibria is established. The cost (payoff) structure distinguishes agents attempting to reach goals efficiently and others wishing to indefinitely avoid “bad” sets (Brihaye et al., 2012).
• Lexicographic objectives generalize priorities over reachability, safety, and more: algorithms reduce optimal strategy synthesis to iterative single-objective game solutions. For a constant number of objectives, problems are in NP ∩ coNP; in general, PSPACE-hard (Chatterjee et al., 2020).
• For multi-vehicle collision avoidance and multi-target visitation, scalable architectures combine pairwise HJ reachability with vehicle clustering, enabling O(N^2) per timestep online complexity and provably safe coordination among dozens of dynamic agents (Shih et al., 2021).

5. Applications and System-Level Properties

Safe-reachability objectives are implemented in domains including but not limited to:

• Autonomous UAV platooning and aggressive flight (Chen et al., 2015, Kousik et al., 2019)
• Real-time manipulator motion planning (Holmes et al., 2020)
• Autonomous racing on ego-vision (Chen et al., 2021)
• Safe navigation in partially unknown or dynamic environments (Bajcsy et al., 2019)
• Data-driven safety verification for LLM-controlled robots (Hafez et al., 5 Mar 2025)
• Latent-space safety for visual manipulation (Nakamura et al., 2 Feb 2025)

Safety/liveness properties are often maintained via a “wrap safety around liveness” design: performance controllers are opportunistically deployed inside the safety envelope, while any impending unsafe condition triggers a fallback to a certified safe controller (Chen et al., 2015, Hsu et al., 2021, Ganai et al., 2023).

The following table highlights a selection of problem classes and solution methods:

Domain	Safe-Reachability Formulation	Principal Solution Methods
Hybrid systems	Closed-set forward reachability (safe over-approximation)	Lattice fixed points, Scott continuity, robust abstraction
POMDPs	Probabilistic goal & safety constraints over beliefs	SMT-based, symbolic constraint search
RL (MDP)	Persistent safety & reward-optimality via violation probability	REF Bellman recursion, Lagrangian RL (RESPO)
Continuous multi-agent	Avoid–reach tubes via HJ PDEs, decentralized guarantees	Hierarchical HJ reachability, clusters, local ILPs
Data-driven/LLM-robotics	Plan-tube safety via Lipschitz zonotope overapproximation	Regression, gradient projection, adjustment loop

6. Robustness, Complexity, and Limitations

Robust safe-reachability ensures that analysis and synthesized policies are insensitive to small perturbations of initial states or model errors. The robust property for a reachability operator is equivalent to Scott continuity on compact state spaces (Moggi et al., 2017). Computation is classically intractable due to the curse of dimensionality for HJ PDEs; advances in neural and data-driven solvers (DeepReach-type) reduce scaling with dimensionality by trading precision for verifiable error bounds (Lin et al., 2022, Nakamura et al., 2 Feb 2025).

Algorithmic guarantees vary:

• Discrete games: finite-memory equilibrium existence (for safe-reachability objectives) is constructive, but computationally EXPTIME in size of the game graph or exponential in the number of objectives (Brihaye et al., 2012, Chatterjee et al., 2020).
• RL and stochastic settings: convergence and constraint satisfaction can be established almost surely under mild conditions (Ganai et al., 2023, Hsu et al., 2021).
• Coverage of hazards in data-driven and latent-space approaches is limited by training distribution; current methods are inherently limited to hazards that can be encoded or induced in world-model imagination (Nakamura et al., 2 Feb 2025, Hafez et al., 5 Mar 2025).
• Real-time receding-horizon approaches guarantee perpetual safety through cycle-by-cycle fail-safe planning, but may introduce conservatism or require fail-safe halts if no feasible plan exists (Kousik et al., 2019, Holmes et al., 2020).

7. Empirical Performance and Practical Evidence

Across robotic and autonomous domains, safe-reachability-based methods have demonstrated:

• Zero observed collisions in hundreds of randomized aggressive flight scenarios when using online FRS/zontopic reach filtering (Kousik et al., 2019)
• Verifiably collision-free manipulator operation in real time across randomized and adversarial environments where CHOMP-based plans resulted in collisions (Holmes et al., 2020)
• In safe RL, up to 3× reduction in constraint violation and up to 50% improvement in reward compared to state-of-the-art baseline methods (Ganai et al., 2023)
• Scalable safe planning for 15+ multi-agent systems with guarantees at O(N^2) per-step complexity (Shih et al., 2021)
• Rapid policy synthesis with incremental SMT solvers for POMDPs, exploring O(10²) plans in spaces with O(10²¹) potential paths (Wang et al., 2018)
• Scenario-based neural PDE correction achieving formal probabilistic safety guarantees with up to 10⁷ validation samples showing negligible or zero violations (Lin et al., 2022)

Empirical and formal results thus support safe-reachability as a tractable and generalizable tool for safety-critical decision making, particularly when combined with robust numerical methods, intelligent abstraction, and, increasingly, data-driven estimation and verification.