‘Hello World!’ from an Alternative Branch of the Tech Tree. This small @circuitmess Wheelson robot (provided by the courtesy of @AlbertGajsak) can learn two opposite behaviours: chase a ball or run away from it. More importantly, it can adapt from one to the other while it is running. All learning happens locally on its ESP32-WROOM-32 module by @EspressifSystem , a conventional, few-dollar microcontroller. There is no GPU, no backpropagation, cloud service, Internet connection, offline retraining or deployment of a new model version. The controller is a spiking neural network (SNN), learning online through biologically inspired, local rules rather than gradient descent and backpropagation. What happens in the video: 1.) A hand-coded teacher observes the detected ball position and indicates whether the robot should steer left, right or go forward. 2.) The SNN learns this steering policy on the device. 3.) I disable the teacher, and the SNN takes control. 4.) I reverse the objective, from chasing the ball to escaping from it. 5.) The teacher is briefly re-enabled, and the same network adapts online. 6.) I then disable it again, leaving the SNN to execute the newly learned behaviour. The task, three-way steering, is deliberately simple. Ball detection is not learned end-to-end: conventional image processing provides the ball position (note the blue boxes in the video on the Calibration screen). What is interesting here is where and how the adaptation happens: inside a tiny embedded device, during execution, without backpropagation. The ESP32 is not neuromorphic hardware so the energy efficiency could be pushed even further. The event-driven SNNs may become much more efficient on hardware designed for them, including digital, mixed-signal or even analog neuromorphic systems. Why are SNNs not already widely adopted? Useful SNN learning remains difficult to scale. The ecosystem is far less mature than mainstream deep learning, temporal spike dynamics are difficult to train, and conventional processors often cannot exploit event-driven sparsity efficiently. Many high-performing SNNs also depend on surrogate-gradient backpropagation or conversion from conventionally trained networks. To scale this, there are unsolved scientific and engineering problems. That is exactly what makes this alternative branch of the technology tree interesting. My next step is to replace the action-directing teacher with exploration and a local reward signal, allowing the robot to discover the steering policy through interaction. I would be interested to hear from others approaching this space from the research, engineering or investment side. Would you like to see whether the autonomous version works? If you are a chip/robot manufacturer, send me some for testing! :) #SpikingNeuralNetworks #NeuromorphicComputing #AdaptiveRobotics #AI #EdgeAI #EmbeddedSystems #SNNs #ESP32 P.S. I coded this in C++ with the help of Kimi K3 model by @Kimi_Moonshot