DeepSeek just released its new V4 Flash Vision model, and I plugged it into our humanoid! This release is exciting because it goes beyond “DeepSeek can now see images.” DeepSeek is combining the agent/reasoning capabilities of V4 Flash with visual understanding and according to its own benchmarks, the new model already comes close to Anthropic’s Opus 4.8 on multimodal agent tasks. That immediately made me wonder: What happens when you give that kind of model the eyes of a real robot? So I gave our robot “RalCox” a simple goal: “Safely approach the person using the laptop.” RalCox, sitting besides me, took a fresh image from its onboard camera and sent it to DeepSeek. In the software screenshot you can actually see the whole experiment: the top-right window is the K1’s live camera view, while the terminal below shows DeepSeek reasoning about what the robot sees. Instead of asking for a normal image description, I asked it to reason about: > the current environment > spatial relationships >obstacles and action-relevant objects > what the robot should do next > what it cannot reliably infer from a single camera image And that’s where it gets interesting: It identified the table and scattered objects as navigation constraints, reasoned about the likely path toward me, suggested cautiously moving closer while inspecting a potential obstacle and explicitly acknowledged uncertainty around depth, hidden obstacles and surface properties. Of course, this is not yet a robot that “understands the world.” But to me it feels like a very first building block toward it: Perception → context → relevance → uncertainty → action reasoning. And there’s another part I find almost equally fascinating: I basically don’t code. I started playing with this only a few hours after the model was released and built the whole test with AI helping me step-by-step connecting the robot camera, writing the Python, debugging the API and changing the reasoning prompts. A few years ago, building even this tiny robotics experiment would probably have required a developer. Today I can have an idea in the afternoon and test it on a humanoid the same evening. Technically, the setup is still very simple: Robot camera → Python agent on the robot → DeepSeek V4 Flash Vision → environment reasoning DeepSeek itself is still running via API, not locally on the robot. Next experiment: let RalCox actively turn its head, capture several viewpoints, and reason across them before deciding what to do. 🤖