EdgeNavMamba: Mamba Optimized Object Detection for Energy Efficient Edge Devices

TL;DR

EdgeNavMamba tackles the challenge of energy-efficient, real-time autonomous navigation on edge devices by fusing a Mamba-based object detector with knowledge distillation to create a compact detector, paired with a PPO-based navigation policy. The two-branch detector (convolution and SSM) is tailored for edge hardware, and distillation preserves teacher-level accuracy while reducing parameters and computation. Experimental results on real edge platforms and simulators show substantial reductions in model size and energy per inference, with navigation success exceeding 90% in MiniWorld across different complexities. The work provides a practical, end-to-end pipeline for sim-to-real RL navigation with efficient perception, enabling robust edge deployment for autonomous systems.

Abstract

Deployment of efficient and accurate Deep Learning models has long been a challenge in autonomous navigation, particularly for real-time applications on resource-constrained edge devices. Edge devices are limited in computing power and memory, making model efficiency and compression essential. In this work, we propose EdgeNavMamba, a reinforcement learning-based framework for goal-directed navigation using an efficient Mamba object detection model. To train and evaluate the detector, we introduce a custom shape detection dataset collected in diverse indoor settings, reflecting visual cues common in real-world navigation. The object detector serves as a pre-processing module, extracting bounding boxes (BBOX) from visual input, which are then passed to an RL policy to control goal-oriented navigation. Experimental results show that the student model achieved a reduction of 67% in size, and up to 73% in energy per inference on edge devices of NVIDIA Jetson Orin Nano and Raspberry Pi 5, while keeping the same performance as the teacher model. EdgeNavMamba also maintains high detection accuracy in MiniWorld and IsaacLab simulators while reducing parameters by 31% compared to the baseline. In the MiniWorld simulator, the navigation policy achieves over 90% success across environments of varying complexity.

Figures (9)

• Figure 1: Edge platforms with onboard Jetson Orin Nano accelerators: (a) the Unitree Go2 robot dog and (b) the Yahboom Rosmaster wheeled robot.
• Figure 2: (a) Reinforcement Learning (RL) diagram, including the interaction with the environment to maximize reward sutton1998introductionmicro2023-mozhgan1, (b) Architecture of Mamba mamba, used for feature extraction and model efficiency, (c) The process of knowledge distillation hinton2015distilling; the teacher model is trained and frozen, the student model is trained based on the teacher features, logits, and ground truth data.
• Figure 3: The proposed architecture of EdgeNavMamba, consisting of two branches of convolution and SSM for feature extraction. The same architecture is used for teacher and student models with a different set of dimensions. Features, then, undergo a detection process, and the bounding boxes are given to the RL model for navigation to the goal.
• Figure 4: The proposed architecture of EdgeNavMamba, consisting of two branches of convolution and SSM (including LiteSS2D) for feature extraction. The same architecture is used for teacher and student models with a different set of dimensions. Features, then, undergo a detection process, and the bounding boxes are given to the RL model for navigation to the goal.
• Figure 5: The architecture of LiteSS2D block, which is in the SSM branch of EdgeNavMamba, following the overall flow of the SS2D proposed by VMamba liu2024vmamba, but with modifications for better efficiency, mentioned in MedMambaLite aalishah2025medmambalite.