Abstract
Modern visual navigation models, while powerful, often suffer from high computational costs that hinder real-time deployment on edge devices. We present DynaNav, a framework that enhances efficiency by dynamically selecting the most relevant image features and neural network layers for a given scene. By skipping unnecessary computations in simple environments, DynaNav achieves a 2.26× reduction in FLOPs and a 42.3% faster inference time compared to state-of-the-art models like ViNT, while maintaining or improving navigation accuracy and providing better model interpretability.
Introduction
Foundation models for robotic navigation, such as ViNT or NoMaD, rely heavily on Transformer architectures to process visual data. These models typically employ a "one-size-fits-all" approach, processing every pixel and activating every layer regardless of scene complexity. This leads to significant hurdles in deployment on power-constrained robots and creates "black box" models that are difficult to interpret. DynaNav addresses these issues by introducing dynamism into the navigation pipeline, asking whether the full power of the model is necessary for every single step.
Methods
DynaNav introduces two primary mechanisms to inject dynamism into existing Transformer-based navigation policies:
- Dynamic Feature Selection: Instead of processing the entire visual feature map, DynaNav uses a trainable "hard feature selector." Utilizing the Gumbel-Softmax trick, the model learns to generate a binary mask that filters out redundant pixels (e.g., sky or uniform walls) and retains only salient regions like paths or obstacles. This creates sparse representations that are significantly cheaper to process.
- Feature-Aware Early Exit: Recognizing that not every decision requires the full depth of a Transformer, DynaNav allows the model to "exit" the decoder early at intermediate layers if the prediction is deemed stable.
- Adaptive Threshold Optimization: To balance accuracy against hardware constraints (latency, memory, and FLOPs), the system uses Bayesian Optimization to search for ideal exit thresholds.
The system utilizes an EfficientNet-B0 backbone for initial feature extraction and is designed as a drop-in improvement for existing Transformer-based navigation models.
Experimental Results
The researchers evaluated DynaNav across diverse real-world benchmarks (Recon, Go-Stanford, SACSoN, and SCAND) and the CARLA autonomous driving simulator.
Efficiency and Performance
Compared to the baseline ViNT model, DynaNav delivered impressive efficiency gains:
- Computation: 2.26× reduction in FLOPs.
- Latency: 42.3% lower inference time.
- Memory: 32.8% lower memory usage.
Despite the reduced compute, navigation performance actually improved across the four public datasets. In CARLA simulations, DynaNav achieved a higher success rate (72.7% in Scene A) than both GNM and ViNT, demonstrating superior path-following capabilities.
Key Insights and Discussion
- Complexity-Aware Compute: The core innovation is that scene complexity is treated as a variable. The model naturally recruits more resources for complex intersections and fewer for simple paths, mimicking biological efficiency.
- Interpretability via Sparsity: The feature masks act as built-in saliency maps. By visualizing which pixels the model retains, researchers can understand exactly what the robot is prioritizing to make its decisions.
- Hardware-Aware Tuning: The use of Bayesian Optimization allows the model to be specifically tuned for the target hardware, whether it is a high-performance GPU or a low-power drone.
Conclusion and Future Work
DynaNav represents a significant step toward efficient, interpretable robotic navigation. While currently focused on the decoder, future work could extend these sparse selection techniques to the CNN encoder for even greater savings. Additionally, integrating threshold optimization directly into the end-to-end training process could further streamline deployment.