Robot Backflips but not Butlers?
Why some robotics demos work, but others will take more work.
Why are we seeing so many dancing robots doing flips but you don’t have armed or legged robots in the workforce, whether at home or in industry?
The answer is about training robots to do intelligent tasks.
In a simulated world, it’s really easy to model what happens with a rigid floor when you walk on it. That means when you train a policy in simulation, it works in the real world with higher fidelity.
But consider what it would take to clear a table after a meal. Maybe you have a translucent cup with some liquid in it. The fluid moving affects how you should hold and pick it up. The translucency on some color of liquid dramatically affects perception. You know how you can squeeze an open 20oz plastic bottle, but can’t squeeze as much if it has never been opened and under pressure? That dynamic shape affects how you pick it up. And you might be dealing with a cold glass or bottle that could have condensation that affects friction. All of this is really complicated physics!
It’s harder to simulate this task. Now apply that logic to trash bags to take out, clothes to fold, tomatoes to chop, or a wiring harness dragged through a car.
You could train a model just like the dancing and the flips, but it would probably fail too often to deploy it.
This is why so many are working on modeling the world to try to get a higher fidelity training environment. Some are building models of the world that let the robot think about the physical tasks to predict what might happen. Others are training end to end policies that have a lot of examples of humans doing the task, and inside the model might be some version of that world model.
A challenge in either case is training data. It might be for a given robot arm and hand you need data from that hardware design to train the robot. Clone Robotics is building human shaped robots in part because we have so much data already about how human hands interact in the world.
The time it takes to have a robot do a task reliably is the limiting factor in getting robots deployed and scaled. Training a new policy for every tiny job just isn’t feasible if the training takes a huge volume of data and months of time.
Some predict a ChatGPT-like moment where this training problem gets solved. What’s exciting is that the hardware for a humanoid is relatively inexpensive, which means anyone that can afford a car will be able to have a capable robot. But not just yet.
At Tango.vc we want to back robotics companies, and this includes those solving for specific verticals or making tools to solve these problems generally.