The 3D Mind of an AGI Robot

A lot of companies are showcasing their humanoid robots these days. Some are merely concepts, others are available now, and the rest are somewhere between.

Engineers program stationery industrial robots to perform specific tasks with millimeter accuracy. But how do you train humanoid robots to move about and interact with objects autonomously? And more specifically, robots with Artificial General Intelligence (AGI) that are expected to learn how to do any mental and physical task a human can do.

To solve this challenge, I asked myself these questions:

1. How does the human subconscious mind plan movement and object interaction before performing a task?

2. How do you implement that same human planning process in humanoid robots?

3. How can robots mimic movement and object interaction by watching humans?

For movement from one location to another, a humanoid robot needs these basic capabilities:

1. Identify or acknowledge a target destination and plan a route, both within line of sight and out of sight

2. Walk on even and uneven surfaces of various grades

2. Go up and down curbs and stairs

3. Anticipate where moving objects will be and avoid contact or collision

4. Stand itself up if it falls down

For object interaction, a humanoid robot needs these basic capabilities:

1. Identify an object and instantiate a 3D version of it within its mind, either from a library of object models or creating the object model in real-time

2. Translate human movement into body and hand pose wireframes so it can map those positions to its own body and hands

3. Watch a human interact with an object while the human explains each movement and interaction step and the reason behind each movement and interaction

4. Identify the interaction touch points on an object or tool interaction points on an object at each step based on the human interaction

5. Understand interactions at each touch point - i.e., grasp, push, lift, release, turn, etc.

6. Replicate the object's ending position and orientation in 3D for each step taken by each human interaction to ensure it matches (or is better than) the real-world end state - i.e., stacked boxes are perfectly aligned with each other

How do you work on this if you don't have a physical robot with these capabilities? Answer: Game engine. It's not ideal, but at least you can work on the solution.

Considering a robot's capabilities and my experience as a human, I've concluded that the human mind subconsciously plans all body movement and object interaction from start to finish before moving. People may disagree with this, but this is my working hypothesis.

So, for example, if you're thinking of walking into the kitchen and opening the refrigerator door, your subconscious previsualizes all body movements required to complete that action, including all the necessary muscle activations to make those movements BEFORE you start any movement. Suppose someone or something gets in your way as you progress through this previsualized motion. In that case, your subconscious uses your current position at the time of interruption as the starting point to re-previsualize how to avoid the obstacle to achieve the desired ending position.

This means that your household humanoid robot, having an accurate 3D map of your house and its contents, should also previsualize its movement and object interactions before moving, just like you would.

If a robot is outdoors and needs to go from location A to B, it picks a point along the route based on its current line of sight and previsualizes movement to that point - and to the next point in its line of sight, and so on.

In addition to previsualizing its own movement and object interactions, it must do this for you, other people, and robots within and outside its line of sight. The ability to previsualize the movement and interactions of itself and others is why I always say that an AGI needs to live in a 3D world.

If you're not yet convinced that humans subconsciously previsualize end-to-end movement and object interaction, consider the concept of muscle memory. An Olympic gymnast has her entire routine previsualized in her subconscious mind, meaning all her muscle activations are coordinated before she even starts her performance. Her body will move exactly as previsualized as long as her mind is perfectly focused and she's not distracted.

Like humans, each robot will build a library of unique movement and object interaction experiences. Whenever the robot encounters a new environment or objects of similar shape and size, it can reference its library to previsualize its actions before taking action.

As required, a robot can instantaneously share its experiences with robots of the same make and model without repeating the learning process. This ability is beneficial when robots need to work in the same environment, cooperatively or at scale.

Humanoid robots are an exciting addition to our world. They can help us with many tasks, and they're even more beneficial when they can learn to perform any physical task a human can do.

If you have not seen it, you can watch my Top 5 Requirements for AGI video here.

#ai #agi #artificialintelligence #artificialgeneralintelligence #robot