CADENCE: Context-Adaptive Depth Estimation for Navigation and Computational Efficiency

Abstract: Autonomous vehicles deployed in remote environments typically rely on embedded processors, compact batteries, and lightweight sensors. These hardware limitations conflict with the need to derive robust representations of the environment, which often requires executing computationally intensive deep neural networks for perception. To address this challenge, we present CADENCE, an adaptive system that dynamically scales the computational complexity of a slimmable monocular depth estimation network in response to navigation needs and environmental context. By closing the loop between perception fidelity and actuation requirements, CADENCE ensures high-precision computing is only used when mission-critical. We conduct evaluations on our released open-source testbed that integrates Microsoft AirSim with an NVIDIA Jetson Orin Nano. As compared to a state-of-the-art static approach, CADENCE decreases sensor acquisitions, power consumption, and inference latency by 9.67%, 16.1%, and 74.8%, respectively. The results demonstrate an overall reduction in energy expenditure by 75.0%, along with an increase in navigation accuracy by 7.43%.

• The paper introduces a slimmable monocular depth estimation network that dynamically scales active channels based on navigation context.
• A unified DQN policy jointly predicts motion and network slimming, achieving up to 75% energy reduction and a 7.43% increase in navigation accuracy.
• Empirical evaluations on an embedded drone platform demonstrate superior runtime efficiency and real-world adaptability compared to static baselines.

CADENCE: Context-Adaptive Depth Estimation for Navigation and Computational Efficiency

System Overview and Motivation

Resource-constrained autonomous vehicles, particularly drones, demand robust perception for navigating unknown terrains but are limited by onboard computational, energy, and sensor constraints. Deep neural networks (DNNs), especially for monocular depth estimation (MDE), are often prohibitive in terms of latency and power usage. CADENCE introduces a context-adaptive autonomy stack that closes the sensing-actuation loop via a slimmable MDE network and a unified navigation-and-adaptation policy, enabling real-time adjustment of perception fidelity based on mission-critical needs and environmental context.

The CADENCE system environment comprises an autonomous drone equipped with an NVIDIA Jetson Orin Nano, GPS, distance sensor, and monocular camera, providing the foundational hardware for in-the-loop evaluations.

Figure 1: The testbed incorporates a drone with embedded computation, GPS, distance sensor, and monocular camera for HIL evaluation.

CADENCE leverages the adaptable width functionality of slimmable networks. By scaling active channels dynamically, the system reserves high-fidelity computation only where navigation demands it and can fully bypass depth estimation in trivial scenarios, thereby maximizing energy efficiency.

Figure 2: CADENCE illustrated as an autonomy stack: raw sensor input drives adaptive perception through the slimmable MDE network followed by context-sensitive navigation-and-adaptation logic.

Perception: Slimmable Monocular Depth Estimation

The MDE component is realized as a slimmable CNN inspired by DGNLNet [hu2021single]. Slimming factors $\boldsymbol{\rho}$ control active channel ratios in hidden layers. When $\rho = 0$, sensor and computation bypasses are triggered, providing maximal efficiency.

The network is optimized with switch batch normalization and weight decay, iterating over all $\rho$ values to stabilize gradients while minimizing $L_1$ loss between predicted and ground truth depth. Empirical evaluation reveals strong correlations between network size and spatial prediction quality.

Figure 3: Example RGB input, ground truth depth, and predicted depth maps for static versus slimmable networks across multiple sizes.

Blurring increases with aggressive slimming, especially at object edges and sparse structures. The slimmable approach outperforms static networks in $R^2$ score at maximal configurations and supports multiple operation profiles for flexible runtime adaptation.

Figure 4: Comparison of test $R^2$ scores for static and slimmable MDE configurations across varying slimming strategies.

Reducing $\alpha$ and $\rho$ yields substantial energy and latency improvements, with diminishing returns as sizes reach minimal configurations.

Navigation-and-Adaptation Policy

Motion control and network slimming are jointly predicted by a unified DQN policy, addressing the intricate interdependency between perception and actuation. The policy ingests recent depth maps and pose information (via GPS/IMU) as a temporal FIFO queue, outputting paired actions $(a, \rho)$.

Training employs double DQN, converging via carefully shaped rewards penalizing inefficiency and distance, with curriculum learning to scale episode difficulty and avoid catastrophic forgetting.

Figure 5: Learning curve during navigation-and-adaptation policy RL training, validation accuracy converges with early stopping.

Evaluations demonstrate that CADENCE's adaptive configurations—predicting both action and slimming—consistently outperform static baselines in energy expenditure and navigation accuracy. Notably, the accuracy improvement holds even under aggressive slimming, contradicting the typical accuracy degradation of static model reduction approaches. Blurring induced by small network configurations may act analogously to Gaussian noise injection, a hypothesis supported by the unified path completion diversity across various policies.

Figure 6: Final evaluation: Energy vs. navigation accuracy for static and adaptive policies, highlighting CADENCE's superior trade-off.

Interpretability and Context Adaptation Insights

CADENCE's adaptation logic demonstrates robust decision-making based on environmental features. Policy-controlled slimming factor predictions correlate strongly with regional obstacle density and trajectory complexity—higher $\rho$ (network width) is allocated to dense urban segments, especially in confined paths, while lower values suffice in open or near-goal areas.

Figure 7: Mean slimming factors spatially mapped, showing denser regions (e.g., housing clusters) require higher network width.

Figure 8: Adaptation parameter $\rho = 0$0 correlates with obstacle density (mean depth values) and proximity to destination.

The policy's rational adaptation is further evidenced by consistent slimming predictions under persistent environmental conditions and complex interactions between obstacle avoidance, navigation frequency, and path length. When obstacles are detected, $\rho = 0$1 is frequently predicted, exploiting prior depth knowledge for collision avoidance and further minimizing sensor acquisitions.

Numerical Results and Comparative Analysis

CADENCE achieves a 75.0% reduction in energy expenditure, a 16.1% reduction in power consumption, and a 74.8% decrease in inference latency relative to static state-of-the-art baselines. Accompanying these efficiency gains, navigation accuracy increases by 7.43%. This is a strong empirical refutation of the accuracy-efficiency trade-off traditionally observed in model reduction methodologies.

The hardware-in-the-loop validation with NVIDIA Jetson Orin Nano and Microsoft AirSim ensures practical relevance and transferability of results to real-world scenarios, extending beyond simulation-only implementations.

Implications and Future Directions

CADENCE's dynamic fidelity scaling establishes a new paradigm for green IoT frameworks, enabling more robust, efficient, and reliable autonomous operation on power-limited platforms. By tightly coupling perception quality to environmental context and navigational demands, it minimizes mission failure risk due to battery depletion, improves flight duration and reaction latency, and supports concurrent operations by freeing power overhead.

Future directions include expanding context features for adaptation parameter prediction, integrating more sophisticated multi-modal sensor fusion, and broadening adaptation logic via neural architecture search or meta-learning. The interpretability findings may inform safer deployments by revealing causal factors in adaptive behavior. Extension to more complex robot tasks, including SLAM, object manipulation, and multi-agent systems, appears feasible given the modularity of the presented autonomy stack.

Conclusion

CADENCE constitutes a context-adaptive perception-and-control stack leveraging slimmable MDE networks and a jointly trained navigation-and-adaptation policy. It rescinds the traditional trade-off between efficiency and reliability: adaptive scaling of network complexity in direct response to navigational context achieves substantial energy savings while improving navigation accuracy. The empirical validation demonstrates practical viability for resource-limited autonomous systems, with broader implications for dynamic runtime efficiency across IoT robotics and embedded AI.