Ok take 1000 robots and make them try to learn from their mistakes.

[Steven Nelson]

Optimus Robots learn from making thousands of mistakes everyday..

https://www.youtube.com/live/EmF4W8NP5CM?is=0qtsJlTDe4ylalqP

[Chris Albertson]

The current problem is that even with a huge budget and hundreds of engineers, Tesla’s robots are not doing anything better than Unitree’s or Boston Dynamics’ robots.  No one is doing anything useful, and this is not for any lack of trying.

I think there is a theoretical brick wall in the way, and incremental improvement will not work.     It is like trying to make a coal-fire steam train break the sound barrier; bigger wheels and better boilers will make it go faster, but breaking that barrier required very different technology, a rocket-powered airplane.  In other words, a revolutionary, not an evolutionary, technology.  LLMs are running into a brick wall.

What Tesla is doing is low-level RL with hardware in the loop.  They seem to be the first to be able to do this in a big way. and it addresses issues others have with sim-to-real. But what is being learned is only perceptual and motor skills; they are improving only the reptile-level brain.

All AI today simply looks at a sequence of tokens and predicts the next token.   It does not understand why it does things; it just does things.     This is the “brick wall that needs to come down.

What’s frustrating is that I think the insight that breaks down this wall will be something simple and obvious.  Like when Issac Newton guessed that gravity might reach all the way to the Moon and the Sun, that one idea changed everything.   I think we are failing to see something that is in plain sight.  

On Jul 8, 2026, at 10:55 AM, Steven Nelson <teamki...@gmail.com> wrote:

Optimus Robots learn from making thousands of mistakes everyday..

https://www.youtube.com/live/EmF4W8NP5CM?is=0qtsJlTDe4ylalqP

--
You received this message because you are subscribed to the Google Groups "HomeBrew Robotics Club" group.
To unsubscribe from this group and stop receiving emails from it, send an email to hbrobotics+...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/CAOSZ7XRBgFL7F3ZbYEhanchJP7uKUFqe3_CVCYSeNh3V3zaQ%2BQ%40mail.gmail.com.

[A J]

In general, it seems like humans and Bots sense and actuate in different ways. We want the Bot to imitate human behavior but it is a different thing. 

[from search engine]

A robot navigates a new forest path by continuously updating a digital 3D map and calculating mathematically optimal routes using sensors. A human navigates by relying on visual intuition, past outdoor experience, and flexible environmental cues like sunlight and landmark.

The Robot's Approach: Absolute Precision

• Sensory Processing: A robot shoots thousands of laser beams (LiDAR) per second to build a point cloud. It knows its exact orientation via an Inertial Measurement Unit (IMU). [1]
• Path Generation: It analyzes the ground grid by grid. If a slope or a rock exceeds its preset mobility limits, the robot marks it as a dead-end and calculates a new geometric trajectory.
• Failure Point: Sudden environmental changes—like shifting shadows, blowing leaves, or thick mud—can confuse its sensors or cause its wheels/legs to lose traction, leading to software freezing or physical immobilization.

The Human Approach: Contextual Intuition

• Heuristic Processing: Humans do not calculate exact distances. Instead, they look for micro-cues like broken branches, worn dirt (indicating a trail), the angle of the sun, or downhill slopes.
• Biomechanical Adaptability: Human ankles and muscles adjust instantly and unconsciously to uneven roots, soft moss, or slippery mud without needing a software update.
• Failure Point: Humans are prone to cognitive biases, easily losing their sense of direction in heavy fog, darkness, or repeating terrain patterns (like dense, identical pine trees).

[Chris Albertson]

On Jul 9, 2026, at 4:16 PM, A J <aj48...@gmail.com> wrote:

In general, it seems like humans and Bots sense and actuate in different ways. We want the Bot to imitate human behavior but it is a different thing. 

The best robots are very close to humans.  Ignore the ROS2 base SLAM stuff. That was a big deal in the 1990s. To day we can robots that use vision and are trained, not programmed.  Tesla and Wamo are very close to what we do. (more below)

[from search engine]

A robot navigates a new forest path by continuously updating a digital 3D map and calculating mathematically optimal routes using sensors. A human navigates by relying on visual intuition, past outdoor experience, and flexible environmental cues like sunlight and landmark.

The Robot's Approach: Absolute Precision

• Sensory Processing: A robot shoots thousands of laser beams (LiDAR) per second to build a point cloud. It knows its exact orientation via an Inertial Measurement Unit (IMU). [1]
• Path Generation: It analyzes the ground grid by grid. If a slope or a rock exceeds its preset mobility limits, the robot marks it as a dead-end and calculates a new geometric trajectory.
• Failure Point: Sudden environmental changes—like shifting shadows, blowing leaves, or thick mud—can confuse its sensors or cause its wheels/legs to lose traction, leading to software freezing or physical immobilization.

The above could be true if the robot’s controller were hand-coded.   But if the robot was trained by RL, then we know there is no internal map, no “if this then that” logic, and no internal route optimizations.     But are there any end-to-end RL-trained robots?

I think Tesla got it close to right.    Last week, I took a 1,300-mile road trip in a new Tesla with the latest software.   I let the car do all the driving.  It used the rigid map optimization only for the high-level routing, and then all of the execution was clearly analog network RL trained.    I think this is “best of both worlds”.     We even were able to dive on one-lane mountain roads that did not have white paint at night, and it handled the case where another car is going the other way. Our car found a way to let the other car by.  This was all done with RL.

So my point is… The robot can be very human if it uses millions of hours of human data to train a model.  I think the search engine results are not quite up to date.

I am not trying to sell Teslas, but I do think it serves as a good example of what the current state of the art is.

BTW:  This book is good if you know little about Reinforcement  Learning.  It is “The Bible” in its field "Reinforcement Learning:

An Introduction”, second edition, Richard S. Sutton and Andrew G. Barto