Introduction
Robotic navigation in dynamic environments like shopping malls or busy sidewalks requires a Map of Dynamics (MoD)—a statistical model of typical human movement. Traditional MoDs, such as CLiFF-map, rely on discrete spatial grids. However, these grid-based approaches suffer from three major flaws: discretization loss at boundaries, data sparsity in unobserved cells, and temporal rigidity, which makes it difficult to capture continuous changes in human behavior throughout the day.
NeMo-map Architecture
To overcome these limitations, we introduce NeMo-map (Neural Implicit Flow Fields), a continuous, neural-based approach that represents the environment as a function $\Phi(x, y, t)$. Instead of discrete lookups, the model uses an Implicit Neural Representation (INR) to query velocity distributions at any coordinate and time.
The architecture incorporates several key technical components:
- Implicit Neural Representation (INR): The robot queries a neural network with a coordinate $(x, y)$ and time $t$ to retrieve local dynamics.
- Hybrid Feature Encoding: This combines a learnable spatial feature grid with a SIREN (Sinusoidal Representation Network) to capture both fine spatial details and periodic temporal patterns, such as daily cycles.
- Semi-Wrapped Gaussian Mixture Model (SWGMM): The network outputs parameters for a mixture model, allowing for multimodal representations (e.g., intersections where people move in multiple directions). The "wrapped" component ensures angular continuity between 0° and 360°.
Experimental Evaluation
NeMo-map was evaluated on two major real-world benchmarks: the ATC Dataset, containing 92 days of tracking data from a Japanese shopping mall, and the ETH/UCY Datasets, which are standard pedestrian benchmarks for various outdoor scenes. The model demonstrates superior generalization by naturally "filling in" sparse regions and implicitly learning environmental topology—such as the location of walls and corridors—simply by observing human flow patterns.
Key Results
NeMo-map significantly outperforms grid-based baselines in statistical accuracy. On the ATC dataset, it achieved a Negative Log-Likelihood (NLL) of 0.775, compared to 1.527 for the best-performing baseline (Online CLiFF-map).
Beyond accuracy, the neural approach is highly efficient. Inference takes approximately 1.3 microseconds, which is roughly 1,000 times faster than the CLiFF-map baseline. In downstream trajectory prediction tasks (60-second horizons), NeMo-map reduced the Final Displacement Error (FDE) to 10.97m, outperforming specialized deep-learning predictors like TUTR and MID by generating more feasible paths that respect environmental constraints.
Conclusion and Future Work
NeMo-map provides a scalable, continuous, and highly accurate framework for spatio-temporal motion mapping. While it excels at modeling periodic patterns, it currently faces limitations with sharp, sudden environmental changes. Future work aims to incorporate online update mechanisms to allow the neural field to adapt in real-time to unexpected behaviors and further refine how the model handles sharp spatial discontinuities like glass walls.
Summary updated