EdgeFlowNet: 100FPS@1W Dense Optical Flow For Tiny Mobile Robots

TL;DR

This work introduces EdgeFlowNet, a lightweight, EdgeTPU-compatible dense optical flow network designed for tiny mobile robots. By combining image chunking to maximize TPU batch throughput with a multi-scale incremental-flow architecture, EdgeFlowNet achieves onboard ~100 Hz inference at ~1.08 W and outperforms prior edge-based methods in both speed and accuracy. The authors validate the approach through real-world quadrotor experiments on static obstacles, unknown gaps, and dynamic obstacles, as well as extensive simulation and zero-shot MPI Sintel comparisons, demonstrating robust performance under SWAP constraints. The results highlight the practical potential of edge-enabled dense optical flow for autonomous navigation in small aerial vehicles and point to hardware-aware design strategies and future dynamic-resolution optimization as fruitful avenues.

Abstract

Optical flow estimation is a critical task for tiny mobile robotics to enable safe and accurate navigation, obstacle avoidance, and other functionalities. However, optical flow estimation on tiny robots is challenging due to limited onboard sensing and computation capabilities. In this paper, we propose EdgeFlowNet , a high-speed, low-latency dense optical flow approach for tiny autonomous mobile robots by harnessing the power of edge computing. We demonstrate the efficacy of our approach by deploying EdgeFlowNet on a tiny quadrotor to perform static obstacle avoidance, flight through unknown gaps and dynamic obstacle dodging. EdgeFlowNet is about 20 faster than the previous state-of-the-art approaches while improving accuracy by over 20% and using only 1.08W of power enabling advanced autonomy on palm-sized tiny mobile robots.

Figures (9)

• Figure 1: Lightweight multiscale EdgeFlowNet architecture for high-speed and accurate dense optical flow estimation.
• Figure 2: Sequence of images of quadrotor navigating through different scenarios: (Top to bottom): Static obstacle avoidance, flight through unknown gaps and dynamic obstacle dodging. The green and red arrow shows the path of the robot and the dynamic obstacle respectively. Inset images at the yellow highlighted trajectory location (top to bottom) show the quadrotor camera view, optical flow and obstacle mask.
• Figure 3: Various representations for navigation (Each row, left to right): RGB, MiDaS depth, Optical flow from RAFT, NanoFlowNet, EdgeFlowNet GPU, EdgeFlowNet GPU with chunking, EdgeFlowNet EdgeTPU with chunking.
• Figure 4: Optical flow predictions for various ways of changing input shapes. Top row: Input images (when resized), middle row: Outputs for resized inputs, last row: Outputs for chunked inputs. Yellow box shows the output at original resolution of $480 \times 352 px.$ Cyan box shows the ground truth. Columns 2 to 5 have the following input sizes: 240 $\times$ 176, 128 $\times$ 96, 64 $\times$ 48, 16 $\times$ 16$px.$
• Figure 5: Maximum speed $V$ variation with perception latency $T_{s}$ for (top) varying robot size $L$ and detection rate $DR$ and (bottom) varying sensing range $Z$ and detection rate $DR$.