Hey friend
I want to tell you about something cool that is happening in humanoid robotics. It is called zero-shot control. This means that a robot can do tasks that it has never been trained to do before.
You can tell a robot like Optimus or Figure 01 to do something like “pick up the cup and hand it to me” or “walk there while carrying this box.”. The robot will just do it. It does not need any training or practice.
So how does the robot do this? It is because of some smart math and algorithms. These algorithms have been getting better and better over time. They started with something called optimal control and now they use something called diffusion policies.
Let me explain this to you in a way.
1. Classical Optimal Control
For a time robots used something called optimal control to move around. This is like a math problem that the robot solves to figure out what to do. The robot uses something called equations to understand how its body moves.
Then the robot solves another math problem to figure out what actions to take. This problem is like: “what are the best actions for the robot to take in the few seconds so that it does not fall over or hurt itself?”
This way of controlling the robot works well when the robot is doing something it has done before.. It does not work as well when the robot is doing something new.
2. Reinforcement Learning and Imitation Learning
To make the robot better at doing things researchers tried something new. They combined two ways of learning: reinforcement learning and imitation learning.
Reinforcement learning is like trial and error. The robot tries something. Sees if it works. If it does the robot gets a reward. If it does not the robot tries something
Imitation learning is like watching someone else do something and then trying to do it yourself. The robot watches a do something and then tries to do it too.
This way of learning made the robot much better at moving and doing things.. The robot still had trouble doing completely new tasks.
3. The Rise of Zero-Shot Capabilities
The big breakthrough came when researchers started training the robot on an amount of data. This data included lots of tasks and scenarios.
Of training the robot to do one specific task researchers trained the robot to do many tasks at once. This way the robot could learn to generalize and do things that it had never done before.
There are two ways that researchers are doing this now:
A. Diffusion Policies
One way is called diffusion policies. This is like a math problem that the robot solves to figure out what to do. The robot starts with an action and then slowly refines it until it is smooth and feasible.
This way of controlling the robot is really good at making the robot move in an human-like way. It can also handle lots of joints and actions at the same time.
B. Transformer-Based Architectures and Optimal Control Priors
Another way is to use something called transformer-based architectures. This is like a neural network that can understand language and vision. The neural network tells the robot what to do. Then the robot uses optimal control to make sure it does it safely and correctly.
Why Zero-Shot Control Is So Hard for Humanoids
Controlling a humanoid robot is really hard. The robot has to understand what to do balance itself and move its joints smoothly at the same time.
This is why zero-shot control is one of the problems in robotics.
My Personal Take
I think we are seeing a change in robotics right now. A years ago every new task required months of training and practice. Now we are moving towards a system where the robot can learn to do things with little or no extra training.
I think the next few years will be, about combining classical robotics math with modern AI techniques. The teams that can do this the best will be the ones that win. They will be the ones that create robots that can really understand and interact with the world.