Tesla Unveils New Abilities For Optimus Robot, Now Walks Better Than Joe Biden! | WLT Report Skip to main content
BoomMag: Readers and Fans
Advertise With Us Newsletter

WLT Report WLT Report

Advertise With Us Newsletter
We may receive compensation from affiliate partners for some links on this site. Read our full Disclosure here.

Tesla Unveils New Abilities For Optimus Robot, Now Walks Better Than Joe Biden!

4.8k views

This is a really incredible update to show you on the Tesla Optimus robot….

I am so excited to be able to buy one of these things and put it in my house!

Yes, I know the robots and AI super-intelligence will eventually kill us all, I get it.  I’ve seen Terminator and T2.  I know Skynet eventually kills us all.

And if it’s not Terminator, you can take your pick of killer robot futures….

ADVERTISEMENT

The Cylons from Battlestar Galactica?

The Borg from ST:TNG?

The Kaylon from The Orville?

I get it.

But until then, I’m going to have a few of these bad boys in my house mowing my lawn, cleaning up the kitchen each day, scrubbing the floors.  It’s going to be incredible.

And there is a chance we avoid the killer robot future and get an army of Data’s from ST:TNG instead, how cool would that be?

I’d love to hang out with Data all day long, that guy was awesome!

Truth is, whether it’s good or bad, that reality is coming quicker than you think.

Elon Musk posted a video showing incredible human walking skills, much less robot like and much more like a chill dude just cruising down the street without a care in the world.

Milan Kovac, who works at Tesla, then reposted this this additional detail:

ADVERTISEMENT

He says “Entirely trained in simulation with RL” and if that feels like a foreign language, allow me to break it down a bit for you so you can understand the stunning reality of what he just said.

Basically, these robots have been trained how to walk like a human entirely in the Matrix.  All simulated.  Doing thousands and thousands of simulations in virtual reality.

The benefit is not falling down and incurring damage in real life, plus the simulations can run much faster and expose the bots to magnitudes more data.

It’s really quite incredible.

Here’s more on how RL works:

What is Reinforcement Learning (RL)?

ADVERTISEMENT

Reinforcement Learning (RL) is a type of machine learning where an agent (in this case, the Optimus robot) learns to make decisions by interacting with an environment. The agent takes actions, receives feedback in the form of rewards or penalties, and adjusts its behavior to maximize the total reward over time. This process mimics a trial-and-error approach, but it’s guided by a mathematical framework.

Here’s how RL works at a high level:

  • Agent: The learner or decision-maker (Optimus robot).

  • Environment: The world the agent interacts with (a simulated space in this case).

  • State: The current situation or configuration of the environment (e.g., the robot’s position, joint angles, or balance).

  • Action: The decision the agent makes (e.g., moving a leg forward or adjusting its arm).

  • Reward: Feedback from the environment based on the action (e.g., a positive reward for a smooth step, a negative one for stumbling).

  • Policy: The strategy the agent develops over time to choose actions that maximize rewards.

In RL, the agent doesn’t need to be told exactly what to do at every step. Instead, it learns through experience by exploring different actions and observing their outcomes. This is particularly useful for tasks like walking, where it’s hard to manually program every possible movement for a robot.

The Wikipedia entry on RL (web ID: 0) explains that RL often uses a Markov Decision Process (MDP) framework, where the agent doesn’t need a perfect mathematical model of the environment—it can learn from sampled experiences, making it ideal for complex, dynamic tasks like robotics.

What is a Simulated Environment?

A simulated environment is a virtual, computer-generated space that mimics the real world (or a simplified version of it) where the agent can interact and learn without real-world consequences. For Optimus, this might be a digital model of a room, factory floor, or even a physics-based simulation that replicates gravity, friction, and other forces.

ADVERTISEMENT

The advantages of using a simulated environment, as noted in the web results, include:

  • Safety: The robot can make mistakes (e.g., fall over) without causing damage or injury. The AI Masterclass (web ID: 4) highlights that RL environments provide a “safe space for AI systems to learn through trial and error,” where real-world mistakes might be “risky, costly, or completely impractical.”

  • Speed and Scale: Simulations can run much faster than real-time, allowing the robot to “experience” thousands of scenarios in a short period. NVIDIA’s Isaac Lab (web ID: 5) emphasizes how GPU-accelerated simulations can process sensory data and physics calculations quickly, streamlining the learning process.

  • Cost-Effectiveness: As mentioned in the AI Masterclass (web ID: 4), simulated environments eliminate the need for extensive manual data labeling, reducing costs compared to supervised learning methods.

  • Control: Researchers can manipulate the environment to test specific scenarios (e.g., uneven terrain, obstacles) or adjust parameters like gravity to challenge the robot.

In the case of Optimus, the simulated environment likely includes a virtual version of the robot’s body, with its joints, sensors, and actuators, interacting with a digital space that replicates the physics of walking—balance, momentum, friction, etc. The X post by Milan Kovac (Thread 1, Post ID: 1907329830443720946) notes that the robot’s walking was “entirely trained in simulation with RL,” meaning all the learning happened in this virtual space before being applied to the real robot.

How Does RL in a Simulated Environment Work for Optimus?

For Optimus to learn how to walk, the process would look something like this:

  1. Setup the Simulation:
    Engineers create a virtual environment using a tool like NVIDIA Isaac Sim (mentioned in web ID: 5). This includes a digital model of Optimus with its physical properties (weight, joint limits, etc.) and a virtual space to move in.

  2. Define the Task and Reward:
    The goal is to walk with a “smoother heel-to-toe gait, straight knees, and arm sway,” as Milan Kovac described. A reward function is designed to give positive feedback for actions that lead to stable, human-like walking (e.g., maintaining balance, moving forward) and negative feedback for undesirable outcomes (e.g., falling, jerky movements).

  3. Trial and Error in Simulation:
    The RL algorithm controls Optimus in the simulation, trying different movements (actions) like adjusting leg angles or shifting weight. Each action changes the robot’s state (e.g., its position or balance), and the environment provides a reward based on how well the action aligns with the goal of walking naturally.

  4. Learning Through Iteration:
    Over many iterations (or “episodes”), the RL algorithm updates its policy—a strategy for choosing actions—based on the rewards it receives. Monte Carlo methods, as described in the Wikipedia entry (web ID: 0), might be used here, where the algorithm averages the outcomes of many simulated experiences to estimate the best actions.

    ADVERTISEMENT

  5. Transfer to the Real World:
    Once the robot has learned a good walking policy in the simulation, that policy is transferred to the physical Optimus robot. The simulation needs to be accurate enough (in terms of physics and dynamics) to ensure the learned behavior works in the real world, a process NVIDIA Isaac Lab (web ID: 5) helps streamline with GPU-accelerated physics simulations.

Why Use RL in a Simulated Environment for Optimus?

  • Complexity of Walking:
    Walking, especially for a humanoid robot, involves coordinating many joints and maintaining balance in a dynamic environment. It’s impractical to manually program every movement, so RL allows the robot to figure out the best way to walk through experience.

  • Rapid Iteration:
    As Elon Musk notes in Thread 0 (Post ID: 1907320983855215002), the “rate of improvement is indeed rapid.” Simulations allow Optimus to try millions of steps in a short time, accelerating the learning process.

  • Scalability:
    The arXiv paper (web ID: 7) mentions that simulations can be run in parallel on multi-core systems, further speeding up training. This is crucial for a project like Optimus, where Tesla aims to deploy a “legion of robots” in 2025, as Musk stated.

  • Adaptability:
    RL in simulation allows Optimus to learn a general skill (walking) that can be adapted to different terrains or tasks, as highlighted by the versatility of RL environments in the AI Masterclass (web ID: 4).

Challenges of RL in a Simulated Environment

While powerful, this approach has some challenges, as noted in the web results:

  • Reward Design:
    Defining the right reward structure is tricky. If the reward function is poorly designed, the robot might learn undesirable behaviors (e.g., an unnatural gait). The AI Masterclass (web ID: 4) notes that “inappropriate reward structure could lead to suboptimal learning or incorrect behavior.”

  • Simulation-to-Reality Gap:
    The simulation might not perfectly match the real world (e.g., differences in friction or sensor noise), which can cause the learned behavior to fail when applied to the physical robot. NVIDIA’s Isaac Lab (web ID: 5) addresses this by using high-fidelity physics simulations to minimize the gap.

  • Computational Resources:
    Complex simulations require significant computing power, especially for high-dimensional tasks like humanoid walking. The arXiv paper (web ID: 7) highlights that while single-core performance is key, multi-core systems can run multiple simulations in parallel to scale up training.

In the Context of Optimus

For Optimus, “reinforcement learning in a simulated environment” means the robot learned to walk by practicing in a virtual world, using RL to iteratively improve its gait through trial and error. The result, as shown in the X post’s image and described by Milan Kovac, is a more natural walk with “straight knees, smoother heel-to-toe gait, and arms sway.” This approach allowed Tesla to train Optimus quickly and safely, leveraging the power of simulation to achieve what would be difficult or dangerous to do with a physical robot alone.

Eagle eye viewers might have noticed in the video that two Optimus robots were charging in the background:

Super cool!

They just walk themselves over to the wall and recharge themselves!

And yes, if this reminds you exactly of the Borg in Star Trek, you’re not alone:

Optimus has come a long way in just a year….

Here is 2024 walk vs. 2025 walk:

ADVERTISEMENT

I love this post referring to the old was as the “I just shit my pants” walk — because let’s be honest, we’ve all done that walk at least once in our life, right?

You think it’s just a fart, but something more comes out because you’re under the weather?

You clench those cheeks together and hold on for dear life as you penguin walk over to the bathroom.

We’ve all done it, don’t pretend like you haven’t….

Heck, Joe Biden did that walk all the time:

Elon then jokingly posted videos comparing the Optimus walking from 2021, 2022 and 2025 saying, ironically: “we might have peaked in 2021”:

Obviously, the joke is that back in 2021, they only first introduced the concept of an Optimus robot and did so with a man in a white suit.

Oh how far we’ve come in four years!


 

Join the conversation!

Please share your thoughts about this article below. We value your opinions, and would love to see you add to the discussion!

log in

Forgot password?

reset password

Back to
log in
Leave a comment
Thanks for sharing!
Thanks for sharing!
We invite you to become a WLT Report insider. Sign up for our free email newsletter, and we'll make sure to keep you in the loop.