The state of AI in AI robots today. Figure's Helix, Physical Intelligence's π0 models, and Google DeepMind's Gemini Robotics all converge on a similar design. They use a vision-language-action model. A slower "thinking" layer, basically a multimodal LLM, understands the request, looks at the scene and plans steps. A separate fast control layer then turns that into motor commands many times a second. You can talk to it, it can break "clean up the kitchen" into steps, and it can notice that a grasp failed and try again. These models are trained mostly via imitation learning from human demonstrations (teleoperation plus first- and third-person video), with reinforcement learning layered on top, mostly in simulation for locomotion and increasingly in the real world for fine-tuning skills. This way the bot learns how to do a task with many variations. Since no production robot AI systems can learn on the fly (yet, more on that below), there are various mechanisms that are used to handle mistakes or novel situations. The slow thinking layer model sees what just happened (the cup slipped, the drawer is stuck) and replans. It adapts through its context window, not via permanent learning. For reasoning-type problems ("the cup isn't where I expected, look in the dishwasher") this works reasonably well, and it will keep improving as the underlying models improve. Bots also log their failures, allowing the company to retrain offline, and push an update in a few weeks or a month or so, the way Tesla does with FSD. That is continual learning of a sort. It just runs at the fleet level on a cycle of weeks instead of inside each robot. For common situations that's a powerful loop, arguably better than per-robot learning, because every robot benefits from every other robot's mistakes. Then there are agentic-like bolt on memory systems ("this family keeps the pans in the left cabinet") and you cover a lot of the personalization need too. But there’s still something missing from this AI stack. Context-based adaptation is good for reasoning and bad for motor skill. A human can pick up an unfamiliar tool, fumble twice, and have it by the third try, because their motor system actually rewires. A current robot model can plan to "try a different grip," but it can't turn those three attempts into a lasting new skill until the next fleet retrain. The same goes for things specific to one site or one body: a stiff drawer in this particular house, a slippery floor, a gripper whose joints have worn and now behave differently. Fleet retraining handles common failures well but handles rare, local ones badly (ask me how frustrated I am about my Tesla continually driving over the same road dip too fast every damn time). Will this lack of on-bot continual learning/adaptation matter? And/or will true on-bot continual learning become possible? These are the trillion dollar questions right now in AI robotics IMHO. I don’t know the answer to either question. I do know that there are several well funded research efforts to solve continual learning and also to reduce computation by many orders of magnitude which would allow large 1T+ reasoning models to run on-bot without cloud dependency. Whether and when these research efforts bear fruit is anyone’s guess, but I’m cautiously optimistic we’ll see massive improvements within 3-5 years in these areas. In the meantime, we’ll keep iterating fleet training faster and faster - if we end up with overnight model refreshes maybe continual learning won’t be needed as much (but I’d still like my Tesla to have a local agentic-like memory to remember that road dip! My car’s suspension will forever be grateful, as will my back).