A robot learning to fold laundry doesn't watch another robot fold laundry. It watches you.
That's the insight driving a fundamental shift in how robots learn tasks - one that removes the most expensive bottleneck in robotics development. Eric Chan at Rhoda AI has been working on what he calls "direct video action models" - systems that train robots by watching internet video of humans doing things, not by collecting terabytes of sensor data from other robots.
The traditional approach was ruinously expensive. You needed to physically demonstrate a task hundreds or thousands of times, recording every motor position, every sensor reading, every joint angle. Want to teach a robot to open different types of doors? That's weeks of manual demonstration across dozens of door types, all carefully logged. The data collection cost more than building the robot.
How Video Action Models Work
Chan's approach sidesteps this entirely. The model watches video of humans performing a task - opening doors, folding clothes, pouring liquid - and learns the underlying action pattern. Not the specific motor commands for that specific robot, but the concept of the action itself. When a robot needs to replicate the task, the model translates that concept into motor commands for its particular hardware.
This works because the internet already contains millions of hours of humans doing things. Recipe videos, how-to guides, manufacturing footage - all of it becomes training data. The model learns that "opening a door" involves approaching a handle, grasping it, applying rotational force, then pulling or pushing. The specific mechanics vary, but the action pattern is consistent.
The practical impact is substantial. A task that previously required 500 manual demonstrations can now be trained with 10-20 videos scraped from the internet. Training time drops from weeks to hours. More importantly, the model generalises better - it's seen hundreds of different people opening thousands of different doors, not just your specific training setup.
What This Means For Development Costs
The cost reduction is dramatic enough to change what's economically viable. Chan notes that data collection used to represent 60-70% of a robotics project's budget. With video action models, that drops to single digits. The expensive part becomes hardware and deployment, not training.
This shifts who can afford to build robots. Small manufacturers, logistics companies, even individual developers can now train systems for specific tasks without needing a robotics lab and a team of PhD students collecting data for months. The barrier to entry just collapsed.
There's a catch, of course. Video action models work well for tasks humans do regularly and film frequently. Opening doors, picking up objects, basic manipulation - all well-covered on YouTube. Highly specialised industrial tasks with no public video record still need traditional data collection. But that's a much smaller set of use cases than most people assume.
The Robotics Data Problem Is Solved
The broader implication is that robots can now learn complex behaviours faster than humans can demonstrate them. A warehouse robot learning to handle packages doesn't need you to demonstrate every possible box size and weight combination. It watches a few thousand delivery videos and extrapolates the rest.
This is the pattern we've seen in other AI domains - foundation models trained on broad datasets outperforming narrow models trained on hand-curated data. The difference is that in robotics, the cost savings are immediate and measurable. Every manual demonstration you don't need to perform saves hours of labour and equipment time.
Chan's work suggests that the expensive part of robotics is shifting from software to hardware. Training is becoming cheap and fast. Manufacturing, deployment, and maintenance remain expensive. That's a different economics entirely - and one that favours production scale over research depth.
For business owners watching these developments, the question is no longer whether robots can learn a task, but whether deploying them makes economic sense. The training cost just stopped being the blocker.