Introduction
Positioning, Navigation, and Timing (PNT) systems are the backbone of modern autonomous technology. While traditional systems like the Global Navigation Satellite System (GNSS) provide unparalleled precision, they often falter in complex, "GNSS-denied" environments such as deep space, underwater, or dense urban canyons. In contrast, biological organisms exhibit remarkable navigational resilience and energy efficiency using specialized neural circuits. This paper explores the synergy between these two paradigms, proposing a fusion framework to transition from rigid "tool-oriented" navigation to resilient "cognition-driven" systems.
The Traditional PNT Paradigm
Traditional machine PNT relies on high-precision mathematical models and absolute coordinate systems. These systems process data from multimodal sensors—such as LiDAR, IMUs, and cameras—through geometric optimization or state-space models to achieve pinpoint accuracy. However, this approach is computationally intensive and brittle in unpredictable or degraded environments where external signals are lost.
The Rise of Brain-inspired PNT
Brain-inspired PNT (BiPNT) leverages neuromorphic engineering to mimic the "cognitive maps" found in biological brains. By utilizing event-driven sensors and neurodynamic models like Spiking Neural Networks (SNNs) or Competitive Attractor Neural Networks (CANNs), these systems achieve high energy efficiency and adaptability. While they currently lack the absolute accuracy of GNSS, they offer a robust alternative for local navigation and spatial experience decoding.
A Four-Layer Fusion Framework
To bridge the gap between precision and resilience, the authors propose a heterogeneous fusion framework across four levels:
- Observation-Level: Combining traditional sensors with event-driven neuromorphic sensors (e.g., event cameras) to maintain data richness under challenging lighting or high-speed motion.
- Capability-Level: Employing AI and dynamic neural networks to bridge raw sensory data with neurodynamic models, enhancing the extraction of spatial cues like place or grid cell activity.
- Decision-Level: Running traditional and BiPNT pipelines in parallel, using ensemble methods like Mixture-of-Experts (MoE) to dynamically weight outputs based on environmental reliability (e.g., trusting GNSS in open skies but leaning on BiPNT in unknown caves).
- Hardware-Level: A distributed architecture pairing von Neumann chips (CPUs/GPUs) for general tasks with neuromorphic chips (like Tianjic) for ultra-low-power algorithmic execution.
Key Insights and Evaluation
The goal of BiPNT is not to perfectly replicate biology but to achieve a "global optimum"—maximizing navigational benefit while minimizing energy cost. Crucially, the authors argue against treating BiPNT as a strictly "GNSS-free" pursuit; GNSS provides the "God's-eye view" and timing synchronization necessary to anchor local cognitive maps.
To measure progress, a new five-metric evaluation framework is proposed:
- Availability: Task success rate in dynamic environments.
- Accuracy: End-to-end positional error relative to the goal.
- Energy Efficiency: Power consumption (Watts) per unit time.
- Efficiency: Ratio of actual path traveled vs. the shortest feasible path.
- Responsiveness: Compute latency and maximum response delay.
Future Outlook
The transition from a "tool" to a "cognition" paradigm requires overcoming significant hardware and software bottlenecks, particularly the lack of mature neuromorphic development toolchains. Future research should focus on dynamic neural networks and the integration of Large Language Models (LLMs) to provide higher-order reasoning, metacognition, and causal inference, pushing unmanned systems toward true artificial general intelligence.
Summary updated