Priya Jaiswal, a leading voice in venture capital and global market analysis, joins us today to dissect the complex world of humanoid robotics. Fresh off the Humanoids Summit in Silicon Valley, the industry finds itself at a fascinating crossroads, caught between the explosive potential fueled by generative AI and the deep-seated skepticism of seasoned investors. We’ll explore the monumental technical hurdles that remain, the strategic momentum China is building with aggressive government backing, and how lessons from the slow rollout of self-driving cars might offer a roadmap for success. Jaiswal will help us understand what it will truly take for these machines to move from lab curiosities to commercially viable assets in our warehouses and homes.
The Humanoids Summit juxtaposed major excitement with deep skepticism. Considering the “very big hill to climb” mentioned by Cosima du Pasquier, what is the single biggest technical hurdle—like dexterity or sense of touch—and what specific steps must researchers take to solve it in the next few years?
That’s the central tension of the entire field right now. While the progress in locomotion is visually impressive, the single greatest hurdle is, without a doubt, fine-motor dexterity integrated with a true sense of touch. It’s one thing to program a robot to pick up a specific, solid object from a designated spot, which we’ve done for decades in manufacturing. It’s another universe of complexity entirely for a robot to approach a cluttered surface, identify a paper cup, and pick it up without crushing it, or to handle a delicate component with precision. The work being done by startups like Haptica Robotics is critical. The next few years must be dedicated to moving beyond rigid grippers. Researchers need to focus on developing soft, sensor-rich end-effectors and the complex AI models that can process that tactile data in real-time, allowing the robot to instinctively adjust its grip pressure and manipulation strategy, much like a human does without thinking.
The article highlights China’s government mandate for a humanoid ecosystem by 2025 and its firms’ dominance at the expo. Beyond government support, what specific technological or supply chain advantages are they leveraging? Could you share a key metric that illustrates the gap between their progress and that of North American companies?
The government mandate is the headline, but the real story is the vertically integrated ecosystem it’s creating. China isn’t just funding research; it’s actively subsidizing the entire supply chain, from high-performance servos and motors to the sensors and processing units. This dramatically lowers the cost of experimentation and production. You could see this on the expo floor, where the most prevalent models, like those from Unitree, were being used by American researchers because they are simply more accessible and affordable. The clearest metric we have right now, according to McKinsey’s analysis, is the sheer number of well-funded players. They’ve counted about 20 humanoid companies in China that have raised significant capital, compared to just 15 in North America. That’s not just a numbers game; it represents a broader, more aggressive, and state-coordinated push that creates a powerful flywheel of innovation and production scale that the West is struggling to match.
Generative AI is credited with jolting the robotics industry. Can you walk me through a step-by-step example of how a visual-language model helps a robot learn a new, complex physical task, and how this differs from the programming methods used just five years ago?
Absolutely. It represents a monumental shift from explicit programming to implicit understanding. Five years ago, if you wanted a robot to clear a table, you’d have to painstakingly code every single motion. You’d define the exact 3D coordinates of a specific type of bottle, program the precise grip force, and map out the exact path to a specific trash bin. If anything changed—a different-shaped bottle, a bin moved six inches—the program would fail.
Today, with a visual-language model, the process is entirely different. First, the robot’s camera scans the scene. You can simply give it a command in natural language, like “Tidy up this workspace.” The model then processes the visual data, identifying objects not as coordinates but as concepts: “laptop,” “coffee mug,” “crumpled napkin.” The “language” part of the model provides the context—laptops are valuable and shouldn’t be thrown away, mugs go to the kitchen, napkins are trash. Finally, the AI translates this understanding into a sequence of actions, generating the motor commands on the fly. It’s the difference between following a pre-written, rigid script and actually understanding the director’s intent.
We see different applications like Disney’s entertainment “Olaf” and Agility’s warehouse “Digit.” What are the different engineering challenges and success metrics for a public-facing entertainment robot versus a task-oriented industrial one? Please provide an anecdote about a specific problem one must solve that the other doesn’t.
The engineering goals are worlds apart. For Disney’s Olaf, the primary success metric is believability and guest safety. Does the robot evoke the character? Does it move in a way that feels organic and charming? The engineering is focused on fluid, expressive animation and, crucially, robust safety protocols for operating in an unpredictable environment filled with children. An anecdote that captures this is imagining a child suddenly running up to hug Olaf. The robot’s systems must instantly detect this, halt its planned path, and react in a way that is both safe and in-character—a monumental software and hardware challenge.
For Agility’s Digit, the metrics are pure industrial efficiency: totes moved per hour, battery life, mean time between failures, and ultimately, return on investment for the warehouse operator like Mercado Libre. Its environment is more structured. The challenge isn’t charming a child; it’s reliably lifting a 35-pound tote thousands of times a day without error. Digit doesn’t need to worry about hugs, but it does need to navigate a semi-structured warehouse floor for an entire shift without needing a human to intervene, which is its own form of complex navigation and endurance problem.
Modar Alaoui compared humanoids to the early years of self-driving cars. Drawing on the eleven-year journey of robotaxis, what is one critical lesson the humanoid industry should learn about public perception or deployment strategy to avoid similar pitfalls? Please elaborate with a specific example.
The most critical lesson is to aggressively manage expectations and pursue a strategy of incremental, proven deployment. The self-driving car industry made a classic Silicon Valley error: it sold a grand, near-future vision of complete autonomy—Level 5 cars everywhere—long before the technology was ready. This led to a cycle of hype and disillusionment. We saw promises of fully autonomous coast-to-coast trips that never materialized, and public trust eroded with every missed deadline and high-profile accident. The eleven-year journey to get Waymo’s robotaxis operating in just a few select cities shows how hard the last 1% of the problem is.
The humanoid industry must avoid this. Instead of promising a general-purpose “Rosie the Robot” that can do your laundry and cook dinner by 2030, they should follow Agility’s model. Focus on one specific, valuable task, like moving totes in a warehouse. Prove its reliability and economic value in that constrained environment first. Build public and investor confidence through demonstrated competence, not dazzling demos. A successful, “boring” deployment in a Texas warehouse will do more to advance the industry than a thousand slick concept videos.
Given the skepticism from pioneers like Rodney Brooks, what specific, measurable performance milestones must a humanoid startup demonstrate in a real-world environment, not a lab, to prove its commercial viability and attract the next major round of investment from previously hesitant VCs?
To win over the skeptics and unlock that next tier of capital, startups have to move beyond the demo and deliver on the economics. VCs, who have historically found robotics to be a capital-intensive black hole, need to see a clear, repeatable path to revenue. First, a startup must demonstrate autonomous operation for an extended period in a real, dynamic environment. For example, a robot would need to successfully perform its primary task—be it stocking shelves or moving boxes—for an entire eight-hour shift in a live retail store or warehouse, without a human operator needing to intervene.
Second, they need to prove the unit economics with a real customer. This means showing that the robot accomplishes a task at a cost-per-action that is competitive with, or superior to, human labor when you factor in the total cost of ownership. Finally, they need to demonstrate adaptability. Show that if a box is knocked over or an aisle is partially blocked, the robot can recognize the anomaly and either navigate around it or alert a human for a different task, rather than just freezing. Hitting these milestones of endurance, economic value, and adaptability is what will turn a cool tech project into a business that a VC can’t afford to ignore.
