Transcription
Humanoid robots have made splashy headlines in the consumer press since Honda's Asimo debuted in 2000, but operational units outside of the lab have remained extremely rare. Tesla's announcement of Optimus in 2021 drew attention, but progress appeared sluggish until the spring of 2025 when two videos were released just 7 days apart. A playful dance on May 14th followed by a home task performance on May 21st; the short interval between the two signals a company eager to demonstrate momentum on both breadth, with the number of tasks, and depth, with autonomy within each task. For investors like us, the footage serves as real-time evidence that Optimus is moving beyond the experimental phase and entering the real-world functionality stage.
So, what exactly has the Tesla bot Gen 3 accomplished? Can we truly place our trust in a robot to handle everyday household tasks proactively and independently? Perhaps even more reliably than a human housekeeper? Welcome to the world of Tesla cars. So, what updates is the Tesla bot Optimus Gen 3 rolling out? Tesla's latest Optimus clip, titled "I'm Not Just Dancing All Day, okay?" Mark's 66-second reel features five distinct household tasks: taking out the trash, brushing crumbs into a dustpan, tearing off a paper towel, and wiping the table, stirring a pot on the stove, and vacuuming the floor. Each segment begins with a voice prompt recorded off camera. There are no visible fiducial markers, QR codes, or scripted guide paths. While some footage is sped up two times, the raw clips still show stable foot placement, compliant arm trajectories, and finger articulation precise enough to separate a single sheet of paper towel.
Whereas earlier demos relied on remote operation or discrete behavior modules, Tesla claims an end-to-end neural network powers the entire sequence. What does that mean? The cost of task switching is going down, a prerequisite for real-world work both at home and on the factory floor. Tesla Vice President Milan Kovich called this latest capability a significant breakthrough in transferring human demonstrations directly into robot policy without the need for extensive teleop fine-tuning. The team's next milestone is to scale data collection to include third-party internet videos, followed by reinforcement learning practice in both simulation and the real world. This kind of pipeline reflects recent academic advances such as diffusion policies for large action spaces and humanlike imitation data sets, treating the internet as a near limitless guidance buffer for embodied AI. If successful, Tesla could drastically reduce the per-task data collection cost, a major bottleneck long sidestepped by rivals like Figure AI and Agility Robotics.
Tesla's latest video opens with what appears to be a simple scene: a soft plastic trash bag resting on a chair. But from the very first movement, it's clear this isn't just a routine mechanical action. Optimus approaches using three fingers to grip the bag's thin edge, then executes a precise 38° wrist rotation, just enough to free the bag from where it clings to its arm and torso. It then shifts its grip to the gathered fabric, redistributing the force vertically and gently places the bag into a swing lid trash can. Quick, smooth, and precise. But the most remarkable aspects lie beneath the surface of this action. Optimus is simultaneously demonstrating three highly sophisticated technical capabilities, each long considered a true barrier in applied robotics: such as material inference. The robot recognizes that the bag is made of a deformable film, flexible yet durable. More importantly, it understands that the material can withstand a strong grip—strong enough, in fact, that it could crush ceramic if misapplied. This reflects real-time physical reasoning and material assessment far beyond the hard-coded routines used by most current robots.
Dynamic grasp planning. While moving, Optimus doesn't hold a single static grip. Instead, it adjusts its grasp, recalculating its hold to keep its center of mass within its support polygon. This is critical for maintaining balance during movement, especially while performing arm tasks midstep. Predictive collision avoidance. As its arm and elbow swing close to the trash can's lid, the robot actively adjusts its trajectory to avoid impact. Not just reacting, but anticipating. This is model-based predictive control, a complex capability typically reserved for high-end robots working in human environments. Why does this matter? Because what you're seeing isn't just a cleaning robot. It's a real-world test of end-to-end policy learning. That means the robot isn't just executing pre-programmed steps, but learning entire behavior sequences from perception to action as a unified process. This is exactly what many academic projects in domestic or service robotics are still trying to model in theory. A small task, but a powerful signal of where Optimus is headed: true autonomy, real-world adaptability, and the ability to make decisions like an intelligent embodied agent.
In the next part of the demo, Optimus is assigned a task that seems simple on the surface: sweeping up breadcrumbs with a brush about 5 inches long and dumping them into a dustpan tilted at a 17° angle. It sounds straightforward, but in robotics, this is one of the toughest tests of force control in real-world environments. Why? Because the interaction between brush bristles and breadcrumbs generates entirely unpredictable feedback forces; the bristles don't move uniformly, the crumbs are irregular and static, and every tiny collision influences the torque on the robot's arm. To handle this, Optimus can't rely on open-loop position control, the simple type that doesn't account for feedback. Instead, it has to modulate force and torque in real time, responding to signals like a living organism reacting on instinct. What makes this possible? The all-new Gen 3 hand. Unlike the previous Gen 2 version, which featured 11 degrees of freedom across the palm and fingers, the Gen 3 hand boasts 22 DOF, making it significantly more compliant, adaptive, and dextrous when dealing with uneven or delicate objects. This is the same hand that stunned viewers back in November 2024 when Tesla demoed it catching a tennis ball, a classic test for fast, precise force control.
And as if to underline its progress in force handling, Optimus follows up with a deceptively delicate move: tearing a single sheet off a perforated paper towel roll. That might sound like a matter of simply pulling hard or soft, but it's not. Pull too slowly and the sheet won't tear along the perforation. Pull too fast or unevenly and the whole roll might unravel, or the sheet could tear off track. Optimus has to modulate its pulling speed just below the tear propagation threshold, solving a micro-level physics problem that's surprisingly complex. This is exactly the kind of development that has robotics researchers excited. It's no longer just about picking up and placing objects. This is real force control; what many consider the gateway for robots to graduate from industrial tasks to flexible domestic work—cleaning, wiping, serving, even cooking. Because only when a robot can control force, not just position, can it begin to interact with the physical world the way humans do. Optimus is proving that's not only possible, it's happening. And if a robot can catch a tennis ball and cleanly tear off a paper towel, then realistically doing your dishes might just be a matter of time.
The moment everyone's been waiting for has arrived. Optimus steps into the kitchen, and it's not just for show. The tomato soup stirring scene is a textbook example. One robotic arm grips the ladle firmly while the elbow rotates in a circular motion, not for flair, but to optimize stirring force based on the resistance of the liquid. As the viscosity increases, the radius of the elbow's motion automatically tightens. The ladle stays neatly inside the pot; no clanking edges, no slips. Meanwhile, Optimus' upper torso leans forward about 6 to 8 cm—a subtle but critical adjustment that keeps its center of pressure stable even as it counters the shifting resistance of the soup. This reveals a central controller that's fusing data from positional sensors, contact force feedback, and depth information from onboard cameras, a form of full-body motion control akin to proprioception in humans. And if you think stirring soup is the main challenge, just wait until Optimus starts vacuuming under the table. This is where every technical layer begins to converge. The robot has to angle the vacuum nozzle toward hidden dust clusters—not just a matter of moving its arm, but a synthesis of SLAM (simultaneous localization and mapping) for spatial navigation combined with rapid centimeter-level wrist height adjustments at the nozzle. In short, it has to know where it is in space and place its hand precisely moment to moment.
And here's the key takeaway: These scenes aren't just choreographed robot ballet. They check off all seven categories of modern robotic manipulation: compression, grasping, tool use, force feedback integration by manual coordination, motion task coupling, continuous workspace coverage. This isn't a robot repeating the same task on a factory line. This is a robot that understands the task, senses the environment, and makes movement decisions like a being with motion awareness.
How will we teach Tesla bot Optimus with video? Tesla is teaching Optimus the way humans learn. And it all starts with video. According to Milan Kovich, the head of the Optimus project, most of the robot's skills don't come from traditional teleoperation-style remote control. Instead, its policy is trained directly from first-person video footage like what you'd get from a GoPro strapped to someone's head. In other words, you wipe down a kitchen counter, the robot watches that video, and then teaches itself how to wipe the counter. No engineers writing code, no labels, no need to track every joint; just raw visual data and a large enough neural network to figure out what's going on. The payoff is huge. First, it slashes the time and cost of supervision. Second, it massively accelerates the range of skills the robot can acquire. And third, it closes the gap between simulation and the real world far faster than competitors still relying on labeled data sets and offline processing pipelines.
According to the latest reports, Optimus is now powered by a single neural network, a unified policy that governs its entire action sequence. That might sound ordinary, but in the robotics world, it's a giant leap. Why does having a single neural network matter so much? Because most robots today still operate using what's called a modular stack, breaking complex tasks into smaller chunks: visual recognition, planning, arm control, force feedback, and so on. Each module is handled by a separate system. It sounds logical, but in practice, these stacks often collapse at the seams where input data or environmental conditions change just enough to throw the whole system off. Tesla, by contrast, is going all-in on a monolithic policy, a single model that learns shared latent representations across many tasks. Instead of stitching modules together, it learns to understand situations in fluid, real-world contexts, just like a human would. While other systems are still trying to teach a robot how to wipe a table by writing thousands of lines of code or manually labeling key points in videos, Tesla is simply letting Optimus watch you work and learn by doing. This is observational learning in its purest form. And it could be the foundation for robots that one day teach themselves to master any task—no coding required.
It all begins with a simple phrase: "Clear the table" or "Stir the pot." For Optimus, each of these commands is the starting point for an integrated chain of actions. A built-in voice recognition system first transcribes the spoken sentence. Then the language is passed into a language-conditioned action encoder which converts the command into a target vector, an actionable signal that guides the robot's entire behavior in context. For decades, the robotics industry has produced incredibly agile, precise, and capable machines. But most domestic robots, even high-profile ones like Honda's Asimo, have failed to gain traction in the market. Not because they lacked physical ability, but because they couldn't communicate in a way humans actually want. Put simply, the problem isn't the robot arm, it's the interaction friction. People don't want to learn a custom GUI or type commands into a digital control panel like they're operating lab equipment. They just want to say something and see it happen. If Optimus can turn a simple sentence into executable behavior, Tesla will have achieved what most service robotics companies are still struggling with: bypassing the custom user interface layer entirely and integrating robots into real life through natural language. That not only makes robots vastly more useful but also slashes the barrier to entry for everyday users—a crucial factor if the goal is to put robots in homes, not just in labs. In a world where people want to say, "Stir the soup," and see it happen without further explanation, Optimus is getting closer than ever to becoming the first robot that truly understands and serves humans, not through code, but through conversation.
If you think holding a ladle, wiping a counter, or vacuuming under a table is just show-and-tell, think again. The home is not a friendly place for robots. It's an adversarial testing ground full of soft, irregularly shaped objects, shifting lighting, pets darting around in a non-stop stream of non-repetitive tasks. Wipe this, grab that, open this lid, place something over there. There's no clear sequence, no rhythmic pace like a factory line. And that's exactly why Tesla's May 21st demo marks a significant milestone. It showed that Optimus has achieved three key technical pillars, three foundational requirements for any robot to move beyond the lab and become a real product. One, manipulation bandwidth. The Gen 3 hand with its 22 degrees of freedom paired with tactile servo control and visual feedback delivers a level of dexterity and force responsiveness high enough to handle soft, slippery, or deformable objects like trash bags or paper towels. Two, contextual autonomy. A single monolithic neural network, no architectural changes needed, can now manage a wide range of unrelated tasks from stirring soup to vacuuming, tearing paper to taking out the trash. No task-specific programming. No decomposing actions into hand-coded steps like in traditional robotics. Three, human-friendly interface. Instead of scripting in an IDE or tweaking technical configs, the user just speaks—"Clear the table"—and the robot listens, understands, and acts. It's the kind of user experience that previous generations of robots simply couldn't deliver.
What about speed? Tesla has been upfront. Some clips in the video are shown at 2x speed, but that's not a weakness. It's a statement of reliability. In a household environment, people don't care if a robot finishes a task in 8 or 11 seconds. They care whether it finishes correctly and completely. How long do you think it'll be before a robot like Optimus shows up in your own kitchen? 3 years, 5 years, or never? If you could teach Optimus just one task using video, what would you want it to learn first? Please share your opinion in the comment section below the video. Thanks for watching our video. If you want to explore more exciting information about Tesla EV or Tesla bot, don't forget to hit the like button and share this video. Also, make sure to subscribe to Tesla Car World and turn on notifications so you never miss our latest videos. We appreciate your support and look forward to seeing you in the next video. Goodbye.