Robots are starting to learn from people, and fleets can share what one robot learns with every other robot. Nobody can yet answer three questions: who taught the fleet this, which robots picked it up and should they have, and how do we take it back everywhere. One Lesson is a governed skill memory for a fleet of home robots, built on MongoDB Atlas. You teach one robot a household skill by talking to it. The lesson reaches six simulated Toronto homes in under a second (382 and 523 ms in our saved runs). Each home's own rules decide whether it takes the lesson: the two homes with kids refuse any lesson that puts a sharp item within a child's reach, and say so out loud. After a bad lesson reaches the homes, ask two homes to "tidy the kitchen" and you watch one robot carry the knife down to a low drawer and the other carry it up to a high cabinet. Ask any robot "put the dishes away" and it answers from what the fleet remembers: "I don't know how to do that yet" before teaching, the dishwasher lesson after. Ask it "who taught you that?" and it says, out loud, who taught it, what that lesson was built on, and how many homes took or refused it. Every lesson carries its teacher, version and family tree. One click rolls back a bad lesson and everything built on it, in one transaction with a receipt. Kill the server, write a lesson to Atlas, restart: every home resumes from its saved change stream token and catches up. The board shows 6 of 6 homes in sync, 0 lessons missing. How we used ElevenLabs Voice is how this fleet's memory is written, questioned and refused. Three ElevenLabs products do it: Scribe to teach and ask, Flash text to speech so six homes each answer in their own voice, and a live Agent that is Home 5. The agent can talk, but it cannot talk its way past the rule: its only powers are three tools that call our code. Scribe (scribe_v2) lets a person teach a robot by talking: "Kitchen tidy. Put the knife and the scissors in the bottom drawer" becomes a lesson, and a deterministic parser turns the transcript into steps, so the safety tags never come from a model. Ask any robot by voice in English, French or Spanish, with Mandarin as a text-checked extra: Scribe detects the language, Atlas finds the English lesson, and the robot answers in that language. Tamil and Hindi did not clear our match floor on the phrasings we tried, so they are not offered. Text to speech (eleven_flash_v2_5) gives each home its own voice: robots say what they learned, what they refused and why, who taught them, and when they rolled back. Home 5's agent has three tools: one checks an instruction against the home's rules and changes nothing, one asks the home to run a lesson, one asks who taught it. Tell it to put the knives in the low drawer and our rule check refuses; the agent explains why. It has no tool to teach, roll back or forget. The demo video's narration is ElevenLabs too.
Team
Awards
