.png)
An AI model is trained on a dataset that captures the world at a single point in time. The physical systems running those models, drones, satellites, humanoid robots, operate in environments that keep changing. A model trained months ago in a data center on data that no longer reflects the field conditions it encounters will degrade over time. Engineers call this model drift. In safety-critical applications, it is a real and growing problem.
The traditional fix is a cloud retraining loop: collect field data, upload it, retrain centrally, push a new model back to the device. In practice, this can take days, requires dedicated engineering overhead, and depends on a reliable internet connection that most field-deployed systems do not have.
On-device AI training addresses this directly: give the device itself the ability to learn continuously, adapting in real time with no connectivity required.
On-device AI training is the process of updating a neural network's parameters directly on the hardware where the model is deployed, without sending data to a cloud server. It is distinct from on-device inference, which runs a trained model to make predictions; training is significantly more computationally demanding than inference.
Three constraints make it genuinely difficult on edge hardware.
1) Power :- The standard optimizers used in AI training are designed to run on digital chips and involve heavy data movement between compute cores and memory. NVIDIA's H100 SXM GPU draws up to 700W during training workloads. Edge devices operate in the milliwatt-to-watt range. Running a conventional training loop on a field-deployed device is physically incompatible with battery-powered operation.
2) Memory :- Training requires storing not just the model weights but also the gradients computed during backpropagation, which are often the same size as the model itself. Edge devices typically have kilobytes or megabytes of RAM, not the gigabytes that conventional training pipelines assume.
3) Connectivity :- Cloud retraining shifts the compute burden off-device but introduces its own constraint: a reliable, low-latency connection. For drones, satellites, and remote industrial sensors, that connection is intermittent at best and unavailable at worst.
Together, these three constraints explain why on-device AI training has remained largely out of reach for deployed hardware, despite years of progress in edge inference.
Vellex has built an analog AI optimization engine. It works by leveraging physical principles like energy minimization that occur naturally in analog circuits, performing mathematical operations that would otherwise require significant digital compute.
Rather than iterating through millions of discrete clock cycles as digital training does, Vellex's hardware settles to a solution through circuit dynamics. The result is real-time AI training at power levels compatible with battery-operated systems, with no cloud dependency.
The practical applications span autonomous robotics, including drones and humanoid systems, satellite systems, and industrial IoT. Drones that adapt their navigation models mid-mission. Satellites that train on-orbit. Humanoid robots and industrial systems that improve incrementally over time rather than degrading between maintenance cycles.
The operational case for on-device training is already visible in the sectors that need it most. Autonomous drones, satellite platforms, and industrial robots are running AI at the edge today on pre-trained, fixed-weight models. Those models degrade in the field, and cloud retraining is operationally expensive and often impractical in these environments. The missing piece has been hardware capable of running training at the required power levels.
The economics reinforce this. As autonomous systems scale across robotics, satellite operations, and industrial deployments, the cost of maintaining model accuracy through cloud retraining scales with them. On-device training breaks that dependency.
For defense and critical infrastructure, the resilience case is equally direct. A system that can adapt in the field without communicating with a remote server is harder to degrade and more operationally robust. It does not depend on a communication link that can be disrupted.
The hardware to make on-device training practical is arriving now. Systems that can adapt without leaving the field will increasingly be the baseline for autonomous robotics, satellite operations, and industrial deployment.
READ MORE
