A Preliminary Exploration of the Differences and Conjunction of Traditional and Brain-Inspired Navigation
Xu He1,2,3, Xiangdong An1,2,3*, Xiaolin Meng1,2,3*, Wenxuan Yin1,2,3, Youdong Zhang1,2,3, Lingfei Mo1,3, Fangwen Yu4, Shuguo Pan1,3, Jingnan Liu5, Yufeng Liu6, Yujia Zhang1,2,3, Wang Gao1,3
- School of Instrument Science and Engineering, Southeast University, Nanjing, China
- The China-UK Research Centre on Intelligent Mobility, Southeast University, Nanjing, China
- State Key Laboratory of Comprehensive PNT Network and Equipment Technology, Southeast University, Nanjing, China
- Center for Brain Inspired Computing Research, Department of Precision Instrument, Tsinghua University, Beijing, China
- Research Center of Satellite Navigation and Positioning Technology, Wuhan University, Wuhan, China
- Institute for Brain and Intelligence, Fudan University, Shanghai, China
* Corresponding authors
Developing universal Positioning, Navigation, and Timing (PNT) remains our enduring goal. This paper asks: How can we endow unmanned systems with brain-inspired spatial cognition while exploiting the high precision of traditional navigation (e.g., GNSS)? In this paper, we provide a new perspective and roadmap for shifting PNT from "tool-oriented" to "cognition-driven" version.
Contributions:
- Multi-level dissection of differences among traditional, biological brain, and Brain-Inspired Navigation (BIN).
- A novel fusion framework integrating key elements of traditional and BIN.
- Forward-looking recommendations for shifting PNT from tool-oriented to cognition-driven.
Frontier Positioning: "Navigation in Complex Unknown Environments" is not merely a grand challenge, but a necessary science & engineering frontier.
PNT Bottlenecks
- GNSS Constrained: denied in deep space, deep ocean, polar regions, and dense urban canyons.
- Dynamic Challenges: motion blur, weak illumination, and confined spaces cripple feature extraction and re-localization.
- Root Deficiency: lack of robustness, adaptability, energy efficiency, and spatial cognition.
Animal Navigation Intelligence
- Humpback Whales: directional error of only ≈1° over hundreds of kilometers during migration.
- Monarch Butterflies: multi-generational migrations spanning hundreds of kilometers.
- Bats: social place cells enabling cooperative spatial navigation.
- Eagles: seamless 3D navigation across aerial, terrestrial, and aquatic domains.
"Brain Kalman" Theory — The brain navigation system and modern navigation systems exhibit isomorphism.
| Cell Type | Functional Analogy | Role |
|---|---|---|
| Place Cells | 🛰️ GPS | Position encoding (Hippocampus) |
| Head-Direction Cells | 🧭 Compass | Directional cues (Parahippocampal) |
| Speed Cells | 📍 Odometer | Motion integration (Entorhinal Cortex) |
| Grid Cells | 🗺️ Path Integrator | Pose estimation & trajectory prediction |
Core Insight — The brain continuously fuses internal and external spatial cues to refine "where I am" and "where I am going."
Irreplaceable Strengths of Traditional Systems
- Ultra-high Precision: GNSS/IMU angular resolution reaches 0.1°; atomic clocks drift only 1 second over hundreds of millions of years.
- Rich Modalities: LiDAR, mmWave radar, RF sensing—beyond biological reach.
- Human Intelligence Extension: real-time global mapping, textual descriptions, big data modeling, and cartographic rules embedded into algorithms.
- Codified Knowledge: human expertise distilled into GNSS and engineering standards.
Key Takeaway — Traditional navigation possesses absolute precision and human-knowledge intervention that biological systems cannot match.
How can unmanned systems simultaneously possess:
- A spatial-cognition-driven "intelligent brain" (BIN);
- The high-precision advantages of traditional navigation?
Ultimate Goal — Emulate the brain, yet surpass the brain.
| Module | Content |
|---|---|
| Differential Dissection | Deconstruct characteristics and differences of different navigation systems (Section II). |
| Fusion Pathway | Propose a reference framework for heterogeneous navigation integration (Section III). |
| Future Perspectives | Outlook for BIN evolution based on the above analyses (Section IV). |
By distilling the essential attributes of each navigation system, this section aims to lay the basis for subsequent heterogeneous navigation fusion while also offering a rapid cognitive outline for the broader researchers.
- Spatial Metric Datum: human-defined coordinate frames (inertial, ECEF, ENU, body, sensor frames).
- Spatial Sensing + Precision Modeling:
- Multi-source Observations: input sources for spatial sensing.
- Precision Modeling: coordinate transformations + hand-crafted rules + state estimation / mathematical optimization.
Note: AI Integration and Its Limitations
- Data-driven methods outperform hand-crafted models in error compensation.
- Remains fundamentally statistical learning—encodes correlations, not causation.
- Fatal weakness: poor out-of-distribution generalization and weak interpretability.
- Compute: MCU / CPU / GPU / AI-dedicated chips (von Neumann architecture).
- Essence: Tool-oriented — relies on precise metrics, hand-crafted rules, and numerical computation. High absolute accuracy, yet fragile when foundations fail.
- Spatial Representation:
- Cognitive maps as the core; no artificial coordinates or numerical concepts.
- External cues + internal states → construct spatial experience.
- Sensorium + Navigation Neural Circuits:
- Multimodal Physiological Perception: vision, audition, olfaction, touch (no ADC conversion; directly generates physiological electrical signals).
- Neural Circuits (Hippocampus-Entorhinal Cortex):
- Egocentric
↔️ Allocentric transformation without numerical rules. - Adaptive regulation through environmental interaction; low energy; no dependency on precise observations (embodied mode).
- Egocentric
- Navigation Capability Spectrum:
- Fundamental: pose estimation, path integration, path planning, etc.
- Advanced: spatiotemporal memory, causal reasoning, strategy optimization, meta-cognition → strong generalization from limited experience.
- Independent Motor Circuits: premotor → primary/secondary motor cortex.
- Essence: Mentality-driven — a closed loop of "perception-cognition-reasoning-decision-control". Highly robust and generalizable, but no absolute precision.
- Spatial Metric: cannot fully escape coordinate frames; Sensor geometry still relies on traditional numerical methods.
- Spatial Sensing + Neurodynamic Spatial Modeling:
- Compatible with traditional & bionic sensors; introduces neuromorphic sensors: event-driven, high dynamic range, ultra-low power.
- Shifting from "hand-crafted numerical rules" to brain-inspired neurodynamic models—mainstream methods: Continuous Attractor Neural Networks (CANNs), Spiking Neural Networks (SNNs), etc.
- Key Advantage: bypasses deterministic models, handles uncertainty, tolerates limited observation precision.
- Trade-off: lower absolute accuracy than traditional methods; The goal is to replicate cognitive mechanisms rather than pursue numerical optimality.
- Navigation Functions (Current):
- Mainstream: pose estimation, SLAM, path integration, planning, decision, control (still within the "perception-planning-control" framework).
- Frontier Minority: complex memory, causal inference, meta-cognition (future focus).
- Compute: introduces neuromorphic hardware (compute-in-memory, high energy efficiency, low latency, low power consumption) while maintaining compatibility with traditional hardware.
- Essence: Paradigm transition — introduces biological cognition and neuromorphic hardware, but has not yet broken out of the instrumental pipeline. A significant gap remains from full mentality-driven navigation intelligence.
| Traditional Sensors (including bionic sensors) | Neuromorphic Sensors | Biological Senses | ||
|---|---|---|---|---|
| Perception Modality | Vision | Mono/Stereo, Depth, Panoramic cameras, Compound eyes, Lidar, etc. | Event cameras | Eye |
| Kinesthetic | IMU, Bionic compass, etc. | — | Vestibular system | |
| Auditory | Conventional microphones, Biomimetic auditory sensors, etc. | Dynamic auditory sensors | Ear | |
| Olfactory | Electrochemical/optical gas sensors, Electronic nose, Bionic olfactory sensors, etc. | Neuromorphic olfactory chip, Neuromorphic E-nose | Nose | |
| Tactile | Capacitive / Piezoelectric tactile sensors, E-skin, etc. | Neuromorphic tactile sensors | Skin | |
| Radio | GNSS / Wi-Fi / UWB / LoRa / Bluetooth, etc. | Neuromorphic radar, Neuromorphic radio frequency sensors | — | |
| Acoustic | Ultrasonic sensors, Sonar, Bionic sonar, etc. | — | Vocal tract | |
| Geomagnetic | Magnetometers, Bionic magnetic sensors, etc. | — | Magnetoreceptors | |
| Communication | 4G/5G, etc. | — | Language, Pheromones | |
| Feature Comparison | Signal Processing | Analog → ADC → Digital | Pulse (event) signals | Electrophysiological signals |
| Data Representation | High-redundancy raw data | Sparse spatiotemporal event data | Neural activities | |
| Power Consumption | Typically ≥ mW | Typically µW | µW | |
| Latency | High (sampling-rate-limited) | Low | Low | |
| Dynamic Range | Limited | High | High |
Key Insight — Neuromorphic sensors enable new perceptual capabilities, yet biological senses still lag behind artificial devices in modality richness and configurability.
| Architecture Design | Memory | Computing Elements | Learning Capability | |
|---|---|---|---|---|
| Traditional Processor | von Neumann-based | Memory and computation separated | Transistors and logic circuits | Software-driven; external training required |
| Neuromorphic Processor | Neuromorphic-computing-based | In-memory / compute-in-memory | Analog neurons and synapses | Synaptic plasticity (e.g., STDP) implemented in hardware; weights adapt dynamically |
Practical Obstacle — Data-protocol incompatibility between neuromorphic and traditional sensors still requires von Neumann intermediaries. End-to-end neuromorphic pipelines remain elusive except for small-scale drones [47].
| Traditional Navigation | Biological Brain Navigation | Brain-Inspired Navigation | |
|---|---|---|---|
| Spatial Metric Datum | Human-defined unification | Spatial experience | Human-defined unification |
| Spatial Sensing Mode | Precise observation | Neural expression | Precise observation (including neuromorphic sensing) |
| Spatial Description | Precision modeling with human-engineered rules | Cognitive maps with spatiotemporal memory | Neurodynamic modeling guided by brain-inspired spatial cognition rules |
| Computational Paradigm | Numerical or learning-based paradigms | Biological neural networks and systems | Neurodynamic simulation paradigms |
| Function Realization | Perception-Planning-Control (tool-oriented pattern, emphasizing high-precision computation) | Perception-Cognition-Reasoning-Decision-Control (mentalization-driven pattern, emphasizing spatial cognition & mental optimization) | > Perception-Cognition-Control (nowadays, tool-oriented pattern) Perception-Cognition-Reasoning-Decision-Control (future, cognition-driven pattern) |
| Hardware | von Neumann-based sensors (including bionic sensors) and processors | Biological sensorium system and the whole brain | Both von Neumann-based and neuromorphic sensors and processors |
| Commonality | All require unified egocentric/allocentric spatial metric transformations, and a usable spatial description model |
- Traditional Navigation: a precision tool—system collapses when foundational assumptions fail.
- Biological Brain: mentality-driven—highly robust yet lacks absolute precision.
- BIN (Current Status): still confined within the instrumental "perception-planning-control" framework; a significant gap remains from full mentality-driven navigation intelligence.
- From "Perception-Planning-Control" → "Perception-Cognition-Reasoning-Decision-Control".
- Break traditional architectural constraints.
- Integrate advanced spatial cognition (memory, reasoning, meta-cognition, etc.).
- Shift from tool-oriented to cognition-driven.
- Fundamental divide: engineering precision vs. brain-inspired cognition.
- Shared bridge: spatial description models.
- Strategy: incremental integration within the All-Source Positioning and Navigation (ASPN) framework; not a radical overturn.
- Two-layer architecture:
- Software layer: heterogeneous fusion & joint optimization across observation, capability, and decision levels.
- Hardware layer: deployment and integration of heterogeneous hardware.
Fusion Keys: the two paradigms are not mutually exclusive.
- Observation-level Fusion: neuromorphic + traditional multi-modal sensors, pushing from "multi-source" toward "all-source" [80-90].
- AI-enhanced perception: TinyML can augment both pipelines, reducing spatial description uncertainty [91].
- Decision-Level Fusion:
- Core tension: BIN robustness/low-latency/energy-efficiency
↔️ traditional absolute accuracy. - Critical open problem: how to formulate a quantifiable, adaptive weighting strategy [101].
- Suggested methods: ensemble learning, imitation learning.
- Implementation: incremental replacement and restructuring of computing modules within the ASPN framework.
- Core tension: BIN robustness/low-latency/energy-efficiency
| Level | Strategy |
|---|---|
| Perception | Heterogeneous sensor configurations for complementary perception. |
| Computation | Tiered co-deployment: MCU/CPU/GPU + AI accelerators + neuromorphic chips (e.g., Tianjic/SpiNNaker2) [69]. |
| Architecture | Distributed processing outperforms centralized aggregation—lower latency, reduced bandwidth dependency. |
| Collaboration | Cloud-edge synergy + V2X heterogeneous navigation architecture (forward-looking). |
- Algorithm deployment: underdeveloped toolchains and ecosystems for neuromorphic chips; von Neumann processors remain necessary alternatives [102].
- Sensor configuration: neuromorphic sensor modalities are still limited; cannot fully replace traditional multi-modal sensing.
- System integration: architectural compatibility, data-protocol alignment, hardware-level time synchronization, and power-performance balancing remain bottlenecks.
Machine Superiority (Single Metrics)
- Sensor sensitivity, resolution, dynamic range, and modality diversity outperform biological organs.
- Configurability and scalability far exceed biology (biological senses lack plasticity after development).
- Even a low-cost IMU achieves 0.1° angular resolution—unmatched by biological brains.
Biological Superiority (Systemic Coordination)
- Natural Unity: no time-synchronization or data-coupling issues across sensors.
- Cognitive Intelligence: general-purpose spatiotemporal generalization, self-organization, self-adaptation.
- Energy-efficiency Miracle: ultra-low power, ultra-long endurance—silicon chips remain far behind.
- Mechanical Agility: environmental adaptability (e.g., a house cat) still exceeds the most advanced robots.
Core Thesis — Machines win on nearly every individual metric, yet their overall navigational competence still falls short of nature. Therefore, both traditional and brain-inspired systems should pursue the global optimum: extracting the greatest navigational benefit at the lowest cost and energy, rather than local precision alone.
Against the "GNSS-Free" Label
- GNSS is a mature, economical, and effective solution; deliberate exclusion is a strategic misstep.
- Humans themselves rely on smartphone location services—traditional navigation and BIN should be complementary partners, not adversaries.
Biological Analogy
- Migratory animals rely on globally invariant cues (stars, geomagnetic field, gravity, polarized light) as a unified spatiotemporal reference [3].
- Insects use pheromones for navigation; GNSS can be viewed as the "pheromone" humanity leaves in geodetic surveying—a public broadcast of spatiotemporal datum and geographic semantics.
Unique Value of GNSS for BIN
- "God's-eye View": compensates for the sub-optimality of brain-localized perception in large-scale navigation [105].
- Timing Dimension Complement: BIN's understanding of temporal coding (time cells, theta-gamma coupling) is limited; GNSS timing capability fills this gap.
- Two conceptual mechanisms:
- GNSS timestamps as temporal anchors for episodic memory.
- GNSS timing driving multiscale time-cell modules to simulate neural oscillator coupling.
Current Bottlenecks: Fragmentation & Stasis
- Existing BIN covers only partial modules (e.g., NeuroSLAM redesigns only the backend; frontend still relies on traditional methods).
- Lacks dynamic modulation capability; cannot replicate biological self-organization and self-adaptation.
Three Key Dynamic Properties of Biological Navigation (Underexplored)
| Dynamic Property | Biological Mechanism | Current Deficiency |
|---|---|---|
| Environment-Adaptive Reorganization | Cell populations reorganize firing frequencies to maintain stable representations across contexts. | Fixed parameters; rigid architectures. |
| Elastic Cognitive Maps | On-demand neuron recruitment and dynamic synaptic adjustment; no need for perpetual full activation. | Lacks elastic computation mechanisms. |
| Task-Dependent Collaboration | Multi-scale coordination and cross-dimensional transitions among navigation cells. | No dynamic modulation models. |
Breakthrough Pathway
- Dynamic neural network technologies (e.g., RNN attractor networks) may provide mechanisms [111, 112].
- Evidence: different learning objectives induce continuous vs. discrete attractor states, enabling task-aware capability switching [113].
Current Problems
- Still in theoretical exploration and partial validation; lacks systematic evaluation frameworks and benchmarks [73, 90].
- Discrete metrics (positioning accuracy, trajectory error) cannot capture overall BIN performance.
Evaluation Philosophy
- Shift from "intermediate localization statistics" to end-to-end navigation capability.
- Core focus: autonomous learning, reasoning, low power, high robustness, and prescribed accuracy.
Five Evaluation Metrics (Current Stage)
| Metric | Definition |
|---|---|
| Availability | Task success rate (%) across complex dynamic environments. |
| Accuracy | Final positional error: distance between actual arrival and intended goal (meters). |
| Energy Efficiency | Power consumed per unit time during navigation (Watts). |
| Efficiency | Ratio of actual path length to pre-computed shortest feasible path. |
| Responsiveness | Algorithmic compute latency: average dynamic latency & maximum response delay. |
Seven-Layer Ecosystem Architecture
- Open-source frameworks & platforms
- Datasets (multi-modal, complex environments)
- Hardware (bionic robots, neuromorphic sensors/chips)
- Standards, specifications & protocols
- Academic ecosystem & talent cultivation
- Privacy & ethics
- Active R&D community network
The future of BIN is not to replace traditional navigation, but to evolve from tool-oriented to cognition-driven—achieving a deep fusion of engineering precision and biological intelligence under the principle of global optimum.
This work is sponsored by the Natural Science Foundation of Jiangsu Province (BK20243064) and a national research grant awarded to Professor Xiaolin Meng as Chair Professor of Intelligent Mobility at Southeast University (SEU), China. Xu He receives support from the Postgraduate Research & Practice Innovation Program of Jiangsu Province (SJCX24_0067) and SEU Innovation Capability Enhancement Plan for Doctoral Students (CXJH_SEU 24204).