Information-Driven Active Perception for k-step Predictive Safety Monitoring

1. Problem Statement

The paper addresses the challenge of predictive safety monitoring in partially observable stochastic systems (e.g., autonomous robots) operating under strict resource constraints (limited bandwidth, sensor energy, or communication budgets).

• The Core Issue: Traditional runtime monitors often assume high-fidelity state observability. In reality, partial observability leads to rapidly diverging predictive uncertainty. If a monitor passively receives intermittent data, it cannot accurately forecast whether the system will enter an unsafe state within a future time horizon ($k$-steps).
• The Goal: Synthesize an active perception policy that dynamically schedules sensor queries. The objective is to minimize the uncertainty (entropy) regarding the safety of future states over a $k$-step horizon while adhering to a limited sensor query budget and minimizing switching costs.
• Formal Model:
  • System Dynamics: Modeled as a Labeled Hidden Markov Model (HMM) with controllable emission functions (the agent chooses which sensor to query).
  • Safety Specification: Defined using Linear Temporal Logic over Finite Traces (LTLf), which is converted into a Deterministic Finite Automaton (DFA) to recognize failure conditions.
  • Product Space: The system and monitor dynamics are coupled into a Product HMM ($Z = S \times Q$), where $S$ is the physical state and $Q$ is the DFA state.
  • Objective: Minimize the average $k$-step conditional entropy of the safety variable $W^k_t$ (a binary variable indicating if a failure occurs within $k$ steps) plus a weighted cost for sensor switching.

2. Methodology

The authors propose an information-driven active perception framework using a policy gradient approach.

A. Information-Theoretic Objective

Instead of minimizing state estimation error, the method minimizes the Shannon conditional entropy of the future safety outcome:
$$H(W^k_t | Y_{0:t}; \pi)$$
where $Y_{0:t}$ is the history of observations and actions. Lower entropy implies the monitor is more confident about whether a failure is imminent.

B. Efficient Computation via Observable Operators

Calculating $k$-step prediction probabilities in HMMs is computationally expensive. The authors utilize Observable Operators to derive an efficient algorithm:

• They define transition matrices ($T$) and observation matrices ($B_\sigma$).
• They construct an observable operator $A_{o|\sigma}$ representing the probability of transitioning and observing simultaneously.
• Proposition 5 provides a closed-form matrix expression to compute the probability of remaining safe for $k$ steps ($P(W^k_t = 0 | y)$) using matrix powers and diagonal projection matrices ($D$) that mask failure states.
• Remark 1: For absorbing failure states (common in safety violations like collisions), the iterative operator simplifies to a single terminal projection, further reducing complexity.

C. Policy Gradient Synthesis

The policy $\pi_\theta$ is parametrized (e.g., using an LSTM) to map observation history to a distribution over sensor queries.

• Gradient Derivation: The authors prove that for a fixed observation history, the conditional safety probability is independent of policy parameters. This allows the gradient of the entropy to be simplified to an expectation form:
  $$\nabla_\theta H(W^k_t | Y; \theta) = \mathbb{E}_{y \sim \pi_\theta} [H(W^k_t | y) \nabla_\theta \log P_\theta(y)]$$
• Optimization: They employ a sample-based gradient descent (using Monte Carlo trajectories) to update parameters $\theta$, balancing the reduction of predictive entropy against the cost of switching sensors.

3. Key Contributions

1. Information-Driven Planning: Introduced a novel framework for predictive runtime monitoring that uses conditional entropy of future safety (rather than current state) as the planning objective in partially observable environments.
2. Efficient Algorithm: Developed a method using observable operators to efficiently compute $k$-step conditional entropy and its gradient, avoiding the intractability of exact dynamic programming.
3. Policy Synthesis: Created a budget-constrained active perception policy that actively schedules sensors to close the gap between partial observability and perfect state information.
4. Empirical Validation: Demonstrated the approach in a dynamic congestion game, showing significant improvements over passive baselines.

4. Experimental Results

The method was validated in a dynamic congestion game involving a robot navigating a 45-node graph with moving obstacles.

• Setup: The robot must transport a reagent from Node 23 to Node 41 while avoiding collisions. It has access to 10 localized sensors with partial coverage.
• Baselines:
  • Uniform Random: Randomly queries sensors.
  • Oracle: Has perfect, cost-free state knowledge (upper bound).
• Metrics: Mean Brier Score (predictive accuracy) and "Gap Closure" (how much the policy closes the gap between Random and Oracle).
• Findings:
  • The trained policy significantly outperformed the random baseline, achieving 30.66% to 68.53% improvement in Brier scores across different horizons ($k=1$ to $15$).
  • The policy closed ~70-74% of the performance gap between the random baseline and the perfect Oracle.
  • Cost-Efficiency: By adjusting the cost weight ($\alpha$), the policy effectively reduced sensor usage (costs) while maintaining high predictive accuracy, demonstrating a successful trade-off between resource conservation and safety assurance.

5. Significance

This work bridges the gap between active perception and formal safety verification.

• Proactive Safety: Unlike traditional monitors that react to data, this method proactively acquires data specifically to reduce uncertainty about future failures.
• Resource Awareness: It provides a rigorous mathematical framework for operating safety-critical systems in bandwidth-constrained or energy-constrained environments (e.g., edge robotics, IoT).
• Scalability: The use of observable operators and policy gradients offers a computationally feasible path to scaling active perception for complex, multi-step safety tasks where exact solutions are impossible.

In summary, the paper presents a robust, information-theoretic approach to making autonomous systems safer by intelligently deciding what to sense and when, ensuring reliable predictive safety monitoring even when full state information is unavailable.