Innovations in Optimus Robot Development

Let’s reverse engineer Tesla Optimus humanoid robot! Part 2: Neural Architecture. Optimus is trained end-to-end: videos in, actions out. I'm quite sure it's implemented by a multimodal Transformer with the following components: (1) Image: some variant of efficient ViT, or simply an old ResNet/EfficientNet backbone (https://lnkd.in/gwJceCJu). The block pick-and-place demo doesn't require sophisticated vision. The spatial feature map from the image backbone can be tokenized easily. (2) Video: two ways. Either flatten the video into a sequence of images and produce tokens independently, or have a video-level tokenizer. There're numerous ways to efficiently process video pixel volumes. You don't necessarily need Transformer backbones, e.g. SlowFast Network (https://lnkd.in/gxZdpeB9) and RubiksNet (https://lnkd.in/gigDhJeT, my paper at ECCV 2020, efficient CUDA shift primitives). (3) Language: it's not clear if Optimus is language prompted. If it is, there needs to be a way to "fuse" the language representations into perception. FiLM is a very lightweight neural network module that serves this purpose (https://lnkd.in/gRskFhwv). You can think of it intuitively as a "cross attention" of language embedding into the image-processing neural pathway. (4) Action tokenization: Optimus needs to convert the continuous motion signals into discrete tokens for the autoregressive Transformer to work. A few ways: - Directly bin the continuous values for each hand joint control. [0, 0.01) -> token #0, [0.01, 0.02) -> token #1, etc. This is straightforward but could be inefficient due to the long sequence length. - The joint movements are highly dependent on each other, which means they occupy a low-dimensional "state space". Apply VQVAE to the motion data to obtain a shorter-length, compressed token set. (5) Putting the above pieces together, we have a Transformer controller that consumes video tokens (optionally with language modulation), and outputs action tokens, one step at a time. The next frame from the table is fed back to the Transformer, so it knows the consequence of its action. That gives the *self-corrective ability* shown in the demo. I believe the architecture is most similar to: - NVIDIA VIMA (my team’s work): https://lnkd.in/gZEDB3fD - Google RT-1: https://lnkd.in/g7N45aCU Lastly, I'm genuinely impressed by the hardware quality. The motions are fluid, and the aesthetics is amazing as well. As I mentioned above, it's such a great decision to follow human morphology closely, so that there is no gap in imitating humans. Atlas from Boston Dynamics only has simple gripper-style hands. In the long run, Optimus' bi-dexterous, 5-finger hands will prove far superior in daily tasks.

I've spent ~4 years working on the problem of locomotion and postural control of humanoid robots. Here is my take on the new Tesla Bot: Optimus Gen 2. There are a lot of little details in the video that the trained eye can unveil: On the positive side: ✅Probably trained using learning from demonstration for the movements. ✅Probably used zero-moment point (a traditional technique for locomotion). ✅The moves are extremely smooth. No video speed-up. ✅It seems the mechatronics improved a lot, there is almost no vibration. ✅10 kg total weight reduction is awesome (higher weight, bigger inertia forces, more difficult to control). ✅Good manipulation (two-hand manipulation of a fragile object is difficult). ✅Probably they use force control instead of position control (the math is more difficult, but you get a smoother control). ✅Exciting that Tesla is pushing robotics forward. Massive kudos. On the negative side: ❌It is still a static robot, as opposed to Boston Dynamics (BD) ones which are dynamic robots. ❌No jumping, no running yet. There is no balance loss at any point (because is damn difficult). ❌Going from static to dynamic movement is so difficult that BD started directly with dynamics. Optimus might need to be redesigned to achieve this. ❌There is very little information about the robot, being more transparent would be good for the industry and the scientific community. Bonus: ✅Let's not forget that Optimus is designed for mass production, this comes with many constraints. ✅Optimus is designed by a product engineer (Elon Musk), as opposed to Boston Dynamics, designed by a researcher (Marc Raibert). If I have to bet who will create a humanoid robot that I can use in my house, I bet on Elon.

AI was phase one. This is phase two. Robots aren’t replacing humans. But they are learning to work with them. Humanoid robots have moved from labs to pilots. Factories. Warehouses. Retail floors. What’s driving it: → Cheaper sensors and AI chips. → Modular designs that scale. → Rising labour shortages. → And leaders bold enough to test early. Who’s deploying right now 👇 Automotive leads: BYD - 1,500 humanoids in 2025, 20k by 2026. Zeekr - UBTECH Walker S1 in swarm trials. BMW - Figure robots in South Carolina. Mercedes-Benz - Apollo robots from Apptronik. Tesla - 5,000 Optimus units for internal use. Logistics & warehousing: Amazon - Digit robots in fulfilment centres. Walmart - Apollo pilots in warehouse sites. Retail & service: Walmart Ghost Kitchens - 240 Richtech units. Chery - 220 humanoids in dealerships. Manufacturing scale-up: Agility Robotics - 10k-unit Oregon plant live. Figure AI - BotQ factory targeting 100k bots. These aren’t PR stunts. They’re early workforce trials. 1–10 robots per site. Still guided and observed. Still learning. Technical reality check: → Reliability remains the biggest gap. → Most humanoids still need human oversight. → Battery life averages 2–4 hours per charge. → The “never stop” dream isn’t there yet. Costs are falling fast. But total ownership still includes integration, downtime, and maintenance. The economics are improving, But not yet revolutionary. Meet the latest line-up 👇 Atlas - agility and motion. Digit - proven in logistics. Phoenix - cognitive AI control. Figure-03 - multi-environment use. Optimus - Tesla’s factory assistant. NEO Gamma - expressive home aid. Apollo - line-ready industrial bot. 4NE-1 - modular forearms. HMND 01 - kitting and handling. Elix - dexterous manipulation. Borg 01 - wheels and legs. A2 - spatial AI vision. AEON - inspection and control. Walker S2 - auto battery swap. LimX Oli - modular SDK. PM01 - open-source research. R1 - acrobatic entry model. Booster T1 - developer platform. Yogi - soft social robot. Abi - empathy-based care. iRonCub MK3 - jet-powered testbed. Strategic truth: This isn't about replacement. It’s about collaboration. Robots take on the repetitive. Humans lead where it matters: judgement, creativity, empathy. Companies winning today treat robots as precision tools, not as people replacements. They design workflows where humans and machines complement each other. That’s the real advantage. Integration, not imitation. The lesson for modern leaders: 1. Ignore the hype (but not the inevitable) 2. Start small 3. Learn fast 4. Blend strengths The future isn’t human or robot. It’s both. Image source: Humanoids Poster 2025 - ver 1.2 - Merphi AB, Sweden. Thanks to Mehrdad Farimani at MERPHI Download the full-res robot poster: https://lnkd.in/eVxcD3p7 🔁 Repost if you're keen to meet the robots. 📥 Try my free newsletter: https://cxo.fm/brief ➕ Follow Rob Llewellyn for more on transformation

🔩 Could a helical electromagnetic motor replace planetary roller screws in humanoid robots? Instead of a conventional motor + Inverted Planetary Roller Screw (IPRS), we can cut the stator and rotor teeth at a helix angle. Rotation directly produces axial motion — eliminating mechanical transmission and its inherent friction losses. Three key advantages over IPRS: ⚡ Near-zero mechanical friction — power losses are shifted from mechanical wear to manageable electromagnetic losses. 🌀 Intrinsic Variable Stiffness (VSA) — the magnetic field acts as a non-linear spring; stiffness is dynamically tunable via current. 🛡️ Passive impact damping — even at zero current, the cogging/restoring force acts as a passive shock absorber, protecting the robot's structure without complex control loops. The Engineering Challenge: Determining the maximum passive axial force (breakaway force) before magnetic "pole slipping" occurs. What’s needed next: → High-fidelity 3D modeling (Ansys Maxwell/Flux 3D) to map the non-linear axial force curves. → Advanced controller development for simultaneous rotary-linear positioning. → Prototype test bench for dynamic load verification. Looking for partners to explore this frontier — researchers, engineers, or companies active in humanoid robotics and high-performance actuators. #Robotics #Actuators #HumanoidRobots #ElectricMotors #VSA #AnsysMaxwell #MotorDesign #Optimus #Innovation

Tesla Reveals Next-Gen Optimus V3 Hand in New Patent: Tendon-Driven with Forearm Actuators & 4 DoF Fingers! Tesla has just published a detailed international patent that appears to reveal the highly advanced hand design for Optimus V3. The system is a sophisticated tendon/cable-driven architecture with actuators located in the forearm (keeping the hand itself lightweight and agile). Each finger offers 4 degrees of freedom, the wrist has 2 degrees of freedom, and the entire mechanism uses just 3 thin, flexible control cables per finger running from forearm actuators through the wrist into the fingers. Advanced wrist routing cleverly switches cables from a lateral stack on the forearm side to a vertical stack on the hand side, with a special transition zone that minimizes stretch, torque, friction, and crosstalk during complex yaw/pitch movements. The design is clearly optimized for mass production with simplified parts and efficient assembly. This patent shows Tesla is solving one of the hardest problems in humanoid robotics — giving Optimus truly human-like dexterity while keeping costs low for high-volume scaling. The future of capable, affordable humanoid robots is getting very real! #Tesla #Optimus #OptimusV3 #TeslaOptimus #HumanoidRobot #ElonMusk #RobotHand #PhysicalAI #TendonDriven #FutureOfRobotics

Tesla has published a new patent detailing a knee joint assembly for Optimus that is directly modeled on human biological anatomy. The patent breaks down how the human knee uses a quadriceps tendon, patella, and patellar ligament system to convert muscle force into powerful bending motion, and then describes how Optimus replicates this using a 4-bar mechanical linkage that mimics the exact same movement pattern. The result is a knee that allows the robot's lower leg to rotate approximately 150 degrees from straight, matching the full range of motion found in a human knee. This enables Optimus to perform natural human movements like walking, squatting, climbing stairs, and kneeling, all of which require the kind of fluid knee articulation that most robotic systems struggle to achieve.

Tesla Optimus — New Generation Revealed by Elon Musk. Elon Musk has unveiled Optimus 3, the newest generation of Tesla’s humanoid robot, signaling a major step toward practical robotics deployment. Unlike earlier prototypes, Optimus 3 is being positioned as a system designed for real-world work, starting inside Tesla’s own factories. According to Musk, the new generation focuses on three critical improvements: 1. Higher Autonomy Optimus 3 can execute more complex, multi-step tasks with reduced human supervision, powered by Tesla’s AI training stack. 2. Improved Dexterity The robot’s hands and manipulation systems are designed to handle precise tasks that were previously difficult for humanoid robots. 3. Real-World Deployment Tesla’s strategy is to deploy Optimus internally first, allowing the robots to learn directly from factory environments before broader commercialization. This approach mirrors Tesla’s playbook in autonomous driving: deploy early, collect massive real-world data, and iterate rapidly. If successful, Tesla could accumulate millions of hours of robotic training data, creating a significant advantage in the emerging humanoid robotics market. The broader implication is clear. Humanoid robots are moving from experimental prototypes to operational tools — potentially transforming industries such as manufacturing, logistics, and services. The key question now is not whether humanoid robots will arrive. It is how quickly companies like Tesla can scale them. #AI #Robotics #Tesla #Optimus3 #Automation #FutureOfWork

Teaching robots to build simulations of themselves allows the robot to detect abnormalities and recover from damage. We naturally visualize and simulate our own movements internally, enhancing mobility, adaptability, and awareness of our environment. Robots have historically been unable to replicate this visualization, relying instead on predefined CAD models and kinematic equations. Free Form Kinematic Self-Model (FFKSM) allows the 𝗿𝗼𝗯𝗼𝘁 𝘁𝗼 𝘀𝗶𝗺𝘂𝗹𝗮𝘁𝗲 𝗶𝘁𝘀𝗲𝗹𝗳: 1) Robots autonomously learn from their morphology, kinematics, and motor control directly from 𝗯𝗿𝗶𝗲𝗳 𝗿𝗮𝘄 𝘃𝗶𝗱𝗲𝗼 𝗱𝗮𝘁𝗮 -> Like humans observing their reflection in a mirror 2) Robots perform precise 3D motion planning tasks 𝘄𝗶𝘁𝗵𝗼𝘂𝘁 𝗽𝗿𝗲𝗱𝗲𝗳𝗶𝗻𝗲𝗱 𝗸𝗶𝗻𝗲𝗺𝗮𝘁𝗶𝗰 𝗲𝗾𝘂𝗮𝘁𝗶𝗼𝗻𝘀 -> Simplifies complex manipulation and navigation tasks 3) Robots 𝗮𝘂𝘁𝗼𝗻𝗼𝗺𝗼𝘂𝘀𝗹𝘆 𝗱𝗲𝘁𝗲𝗰𝘁 morphological changes or damage and rapidly recover by retraining with new visual feedback -> Significantly enhances resilience. The model is also 𝗵𝗶𝗴𝗵𝗹𝘆 𝗲𝗳𝗳𝗶𝗰𝗶𝗲𝗻𝘁, requiring minimal memory resources of just 333kB, making it broadly applicable for resource constrained robotic systems. 𝗧𝗵𝗶𝘀 𝗶𝘀 𝗮𝗹𝘀𝗼 𝘁𝗵𝗲 𝗳𝗶𝗿𝘀𝘁 𝗺𝗼𝗱𝗲𝗹 𝘁𝗼 𝗮𝗰𝗵𝗶𝗲𝘃𝗲 𝘀𝘂𝗰𝗵 𝗰𝗼𝗺𝗽𝗿𝗲𝗵𝗲𝗻𝘀𝗶𝘃𝗲 𝘀𝗲𝗹𝗳-𝘀𝗶𝗺𝘂𝗹𝗮𝘁𝗶𝗼𝗻 𝘂𝘀𝗶𝗻𝗴 𝗼𝗻𝗹𝘆 𝟮𝗗 𝗥𝗚𝗕 𝗶𝗺𝗮𝗴𝗲𝘀, 𝗲𝗹𝗶𝗺𝗶𝗻𝗮𝘁𝗶𝗻𝗴 𝗰𝗼𝗺𝗽𝗹𝗲𝘅 𝗱𝗲𝗽𝘁𝗵-𝗰𝗮𝗺𝗲𝗿𝗮 𝘀𝗲𝘁𝘂𝗽𝘀 𝗮𝗻𝗱 𝗶𝗻𝘁𝗿𝗶𝗰𝗮𝘁𝗲 𝗰𝗮𝗹𝗶𝗯𝗿𝗮𝘁𝗶𝗼𝗻𝘀. I believe the next phase of robotic automation inevitably comes with self-awareness of robots. Self-reflection is a major part of how we as humans improve upon ourselves; as 'general purpose robots' emerge, so would their self-reflection. This enables robots to continuously monitor and update their internal models, thereby refining their performance in real time. This is a huge step towards robot self-awareness! Congratulations to Yuhang Hu, Jiong Lin, and Hod Lipson on this impressive advancement! Paper link: https://lnkd.in/gJ-bkU8N I post the latest and interesting developments in robotics—𝗳𝗼𝗹𝗹𝗼𝘄 𝗺𝗲 𝘁𝗼 𝘀𝘁𝗮𝘆 𝘂𝗽𝗱𝗮𝘁𝗲𝗱!

Tesla Optimus Jogging Isn’t the Story — Cost Control Is Tesla’s latest Optimus demo shows the robot jogging smoothly. Balanced. Confident. Internet goes wild. But jogging isn’t the breakthrough. Here’s what actually matters, based on Tesla disclosures, supplier patterns, and how Tesla scales hardware: • Jogging signals closed-loop control maturity, not a new capability   • The real unlock is actuator cost, not gait elegance   • Tesla is designing Optimus around in-house motors, drives, and electronics   • Target economics matter more than tricks: <$20K long-term BOM vs six-figure peers   • Training leverage comes from Tesla’s autonomy stack, simulation, and data pipelines   • Manufacturing discipline beats one-off demos every time  The bigger signal: Humanoid robots won’t win on viral movement. They’ll win on cost curves, supply chains, and factory throughput. Jogging proves feasibility. BOM control proves scalability. Tesla understands this better than anyone because they’ve already done it with cars, batteries, and autonomy hardware. The race isn’t who looks most human. It’s who can ship millions without collapsing margins. More on robotics, automotive software, AI infrastructure, and scaling realities at SDVGuru.com #Tesla #Optimus #HumanoidRobots #Robotics #EmbodiedAI #Manufacturing #Innovation

𝗜𝗻𝗱𝘂𝘀𝘁𝗿𝗶𝗮𝗹 𝗵𝘂𝗺𝗮𝗻𝗼𝗶𝗱𝘀 𝗮𝗿𝗲 𝗴𝗲𝘁𝘁𝗶𝗻𝗴 𝗶𝗻𝘁𝗲𝗿𝗲𝘀𝘁𝗶𝗻𝗴. 𝗡𝗼𝘁 𝗯𝗲𝗰𝗮𝘂𝘀𝗲 𝘁𝗵𝗲𝘆 𝗹𝗼𝗼𝗸 𝗵𝘂𝗺𝗮𝗻, 𝗯𝘂𝘁 𝗯𝗲𝗰𝗮𝘂𝘀𝗲 𝗳𝗮𝗰𝘁𝗼𝗿𝗶𝗲𝘀 𝗮𝗿𝗲 𝗮𝗹𝗿𝗲𝗮𝗱𝘆 𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲𝗱 𝗳𝗼𝗿 𝗮𝘂𝘁𝗼𝗺𝗮𝘁𝗶𝗼𝗻. Consumer humanoids must handle open-ended chaos. Industrial humanoids get stations, fixtures, takt times, and SOPs — a much narrower software problem. And one company has a structural advantage nobody else can match: 𝗧𝗲𝘀𝗹𝗮 ’𝘀 𝗢𝗽𝘁𝗶𝗺𝘂𝘀. Because Optimus trains inside Tesla’s own factories, it gets immediate deployment environments, continuous data and fast iteration, zero negotiation with integrators, alignment with real production bottlenecks. And soon, a boost from Grok for perception and control. That tight feedback loop is a moat — and likely gives Optimus the fastest time-to-market in the category. The rest of the field has friction: manufacturers, integrators, safety approvals, IT, pilot cycles. Access slows everything. There is one player that might be closing the gap: 𝗔𝗴𝗶𝗹𝗲 𝗥𝗼𝗯𝗼𝘁𝘀. Today’s acquisition of ThyssenKrupp Automotive Engineering gives them something precious: Real car factory access, engineering integration, and validation environments. This arguably puts Agile in the #2 slot for industrial humanoids — not on hardware, but on access. 𝘈𝘯𝘥 𝘪𝘯 𝘩𝘶𝘮𝘢𝘯𝘰𝘪𝘥 𝘳𝘰𝘣𝘰𝘵𝘪𝘤𝘴, 𝘢𝘤𝘤𝘦𝘴𝘴 𝘪𝘴 𝘢 𝘬𝘦𝘺 𝘢𝘥𝘷𝘢𝘯𝘵𝘢𝘨𝘦 The next 12–24 months will show whether these robots move from hype to a real industrial platform. My bet: The winners will be the ones with the tightest data loops, not the most elegant mechatronics. Tesla Grok #Optimus Agile Robots SE thyssenkrupp Automation Engineering