Robots are learning to do housework from videos of humans doing chores

The Quest for the “Robot Internet”: Solving the Data Scarcity Problem

In the world of AI, “scaling laws” are the golden rule. We have seen that large language models and image generators become exponentially more capable simply by feeding them more data. Now, researchers are betting that the same logic applies to robotics. If we can provide enough examples of physical tasks, robots will eventually master the complexities of the human home.

However, there is a massive hurdle: there is no “internet for robot data.” While chatbots can scrape billions of words from the web, there is no central repository of how to fold a shirt or scrub a dish. This has created a desperate race to gather “physical AI” data from the real world.

From Delivery to Data: DoorDash’s Novel Gig

One of the most surprising players in the robot revolution is DoorDash. Known for food delivery, the company has entered the “cottage industry” of data collection. DoorDash is now paying gig workers up to $25 an hour to record themselves performing mundane household chores.

These workers wear head-mounted smartphones to capture first-person perspectives of tasks like folding clothes or washing dishes. This footage is then processed to track the precise movements of the head, hands, and fingers, providing the raw material needed for machine learning algorithms to understand human dexterity.

The Hierarchy of Learning: Video vs. Teleoperation

Not all data is created equal. Experts categorize robot training into different tiers of quality and cost:

1. Human Video Data (The Baseline)

Cheap and abundant, videos of humans performing tasks provide a general understanding of a goal. However, they lack the specific robotic movement commands needed for a machine to execute the task.

2. Robot Teleoperation (The Gold Standard)

As noted by Simar Kareer, a robotics researcher at Georgia Tech, teleoperation—where a human manually operates a robot—is the highest quality data because it includes actual motion commands. The downside? This proves incredibly expensive and slow, as humans move much slower when controlling a robot than when using their own hands.

3. The Hybrid Approach

The emerging trend is to use a large volume of cheap human video data to create a baseline understanding, which is then refined using a smaller, high-cost pool of teleoperation data to teach specific robotic actions.

The Road to Autonomous Home Help

Beyond manual recording, several other trends are accelerating the path toward robots that can handle your laundry:

• Simulated Environments: Some companies are letting robots learn in video-game-like simulations before transferring that software into physical hardware.
• Handheld Grippers: Tools that allow humans to demonstrate tasks using a handheld version of a robot gripper, making the data easier to translate into software.
• Humanoid Ambitions: Figures like Tesla CEO Elon Musk are pushing for robots smart enough to handle home chores, leveraging the same autonomy goals seen in Tesla’s FSD (Full Self-Driving) programs.

While the potential is vast, the timeline remains a subject of debate. Roboticist Ken Goldberg of UC Berkeley suggests that a robot capable of doing your laundry could arrive in two, five, or even twenty years—or longer.

FAQ: The Future of Household Robotics