Learning Robots for Messy Homes

The key change is that home robotics is shifting from robots that follow fixed if then rules to robots that build a live model of the room and decide what to do next. A Roomba can succeed because vacuuming is one repeated motion on the floor. A general home robot has to notice a toy where a shoe used to be, understand a spoken task, plan several arm and body movements, and recover when the scene changes mid task.

• Earlier consumer robots won by narrowing the job until the environment was almost pre solved. iRobot said it passed 40 million robots sold in 2021, which shows how well floor cleaning fit a bounded workflow. That same approach breaks once the robot has to handle cabinets, counters, laundry, stairs, and unfamiliar objects.
  4 irobot 1 sacra
• What changed after 2022 is the arrival of end to end action models trained on large amounts of sensory data. In practice, that means a robot can take camera input plus a task like load the dishwasher, infer where dishes are, choose grasp points, and adjust if one plate is blocked instead of stopping on an unhandled edge case.
  2 sacra 1 sacra
• The bottleneck now is less the motor or the camera and more the data loop. Teams like 1X, Figure, and Foundation are using teleoperation, human demonstrations, and simulation to capture thousands of examples of messy real world behavior, because indoor autonomy has to be learned from many homes, not programmed once in advance.
  1 sacra 2 sacra 3 sacra

The next phase is a split market. Narrow home robots will reach scale first by doing one annoying chore cheaply and reliably, while humanoids use early deployments to collect data until they can string together many chores in the same home. The winners will be the companies that turn every task attempt, including mistakes and human interventions, into better household autonomy.