A new open-source smart glasses platform achieves over 11 hours of continuous on-device AI processing from a 200 milliamp-hour battery, solving one of wearable AI's biggest challenges: power consumption. The breakthrough combines event-based vision sensors, which only record changes in a scene rather than full video frames, with specialized power management that keeps the main processor asleep between inferences.

Why Does Battery Life Matter So Much for Smart Glasses?

Smart glasses face a unique engineering problem. They need to run artificial intelligence models locally on the device to preserve privacy, reduce latency, and enable real-time decision-making, but they must fit in a form factor small enough to wear on your face all day. Traditional cameras consume enormous amounts of power because they capture full-resolution images at fixed intervals, even when nothing interesting is happening. An event-based camera, by contrast, works like the human eye: it only "wakes up" when something changes in the scene, emitting sparse data only during motion or lighting shifts.

The research team, working with an open-hardware platform called OpenGlass, integrated a Prophesee GENX320 event-based camera alongside a standard frame-based imager in a compact form factor similar to Ray-Ban Stories. The system uses a flexible printed circuit interposer that decouples sensor selection from the main circuit board, allowing researchers to swap cameras and algorithms without redesigning the entire device.

How Does the Power Management System Work?

The key innovation lies in the hardware-software co-design. The system combines a configurable power management integrated circuit (PMIC) with an event-driven wake-up mechanism using an nRF5340 coordinator chip. This means the main processor, a GAP9 RISC-V SoC (system-on-chip), stays powered down between inferences and only activates when the event camera detects something worth processing. The result: up to 11.5 hours of continuous on-device machine learning from a 200 milliamp-hour battery, small enough to fit inside eyeglasses.

To demonstrate the system's real-world capability, the researchers tested an egocentric hand gesture recognition pipeline on the LynX dataset, a collection of gesture videos captured in extended reality scenarios. The R(2+1)D convolutional network, a spatiotemporal architecture adapted for event-histogram inputs, achieved 83.94% accuracy with a macro F1 score of 0.781 under strict leave-two-subjects-out cross-subject validation. End-to-end inference latency was 78.3 milliseconds, fast enough for responsive interaction.

Steps to Understand Event-Based Vision for Wearables

• Traditional Frame Cameras: Capture full-resolution images at fixed intervals (e.g., 30 frames per second), consuming power continuously even when the scene is static and nothing changes.
• Event-Based Cameras: Respond asynchronously to luminance changes with microsecond-resolution data, emitting sparse events only when motion or lighting shifts occur, dramatically reducing power draw during static conditions.
• Temporal Histograms: Event streams are accumulated into polarity-separated temporal histograms, a representation that standard video neural networks can process efficiently on microcontroller-class hardware.
• Stochastic Augmentation: Temporal and polarity augmentation strategies during training improved accuracy by 8.9 percentage points, demonstrating that video architectures can be effectively adapted for event-based egocentric gesture recognition.

What Does This Mean for the Future of Wearable AI?

The breakthrough addresses a critical bottleneck in wearable AI development. Smart glasses have the potential to become truly always-on AI companions, continuously sensing the user's environment and providing real-time assistance. However, this vision has been constrained by the power demands of running neural networks on tiny devices. By combining event-based sensing with co-designed power management, the OpenGlass platform demonstrates that full-day wearable operation is now feasible.

The researchers released all hardware designs, firmware, and trained models as open source, enabling rapid prototyping and experimentation across the research and startup communities. This openness is particularly important because it allows the next generation of wearable AI developers to build on proven power-efficient architectures rather than starting from scratch.

Beyond smart glasses, the broader industry is recognizing the importance of edge AI. Synaptics recently unveiled the Astra SRW1500, a single-chip AI microcontroller that combines an Arm Cortex-M52 processor with an Arm Ethos-U55 neural processing unit (NPU), delivering up to 50 GOPS (giga operations per second) of processing performance alongside Wi-Fi 7 connectivity. The company positioned this as a critical enabler for next-generation smart home devices, industrial sensors, and connected products that can make decisions locally rather than relying entirely on cloud-based AI processing.

"IoT is evolving from data-gathering to a continuous intelligence loop, where devices sense and respond in real time. The SRW1500 AI microcontroller draws on the power of Wi-Fi 7 to deliver both the ultra-responsive connectivity and fast inferencing needed to deliver next-generation AI services, for example within the home," said Ananda Roy, Senior Product Line Manager at Synaptics.

Ananda Roy, Senior Product Line Manager at Synaptics

The shift toward on-device inference reflects a broader recognition that running AI locally offers significant advantages: reduced latency, improved privacy, lower bandwidth requirements, and real-time decision-making capabilities. While many AI workloads continue to run in data centers, a growing number of technology vendors see edge AI as a powerful complement to cloud-centric architectures.

The convergence of these trends, from ultra-low-power event-based vision to integrated AI microcontrollers with wireless connectivity, suggests that the next wave of AI-enabled devices will be fundamentally different from today's cloud-dependent systems. Wearables like smart glasses will finally have the power efficiency and processing capability to operate as true always-on AI companions, while smart home devices and industrial sensors will gain the autonomy to make intelligent decisions without constant cloud communication.