Humanoid robots are moving beyond hype-driven prototypes toward early commercial deployment, with automotive manufacturing emerging as the first scalable adoption market, according to IDTechX’s recent report Humanoid Robots 2026-2036: Technologies, Markets, and Opportunities.
According to the report, logistics and warehousing are expected to follow as cost declines and performance improves, while home-use remains a longer-term strategic demand driver. With continued progress in embodied AI and hardware cost reduction, IDTechEx forecasts the humanoid robot market will reach ~$29.5 billion by 2036.
Recent advances in vision-language models and simulation are expanding the ability of humanoid robots to function in real-world settings. Closing the sim to real gap — the mismatch between robot behavior learned in simulation and performance in the physical world — is especially critical for humanoid robots. “The goal of closing the sim-to-real gap is ensuring that the complex reasoning and physical execution learned in a digital space translates into production-grade reliability when the robot is put to work,” said Pras Velagapudi, CTO, Agility.
Velagapudi joined Agility Robotics in 2023, bringing decades of experience in robotics planning and control, creating and deploying robots into home, industrial, and outdoor environments globally. In a landscape constantly disrupted by new technology, he is focused on pragmatic product-relevant innovation: driving vision and strategy for Agility’s Digit, operationalizing AI R&D and exploring what’s new. In this interview, Velagpudi discusses the challenges in closing the sim-to-real gap, what makes a humanoid robot ready for real-world deployment and how far is that threshold today.
Tech Briefs: The physical AI industry wants to close the sim-to-real gap — the challenge of making virtual environments realistic enough that robots trained inside them can operate reliably in the physical world. Why is closing this gap essential for humanoid robots?
Pras Velagapudi: Training a robot to be reliable means spending a lot of time allowing it to explore all the possible things that could happen, and allowing it to learn from that exploration. You need to not only demonstrate how things work when everything is going right, but also how to get things back on track when things aren’t going right: If a foot is starting to slip, or an object is starting to slide out of a grasp. This sort of training is very hard to do in the real world alone, because the real world requires real robots, real environments, and most importantly, real time. This doesn’t scale well for very dynamic motions — there’s simply not enough time to explore all the different physical effects that could happen. For motions like walking, balancing, and whole-body coordination, humanoid robots like Digit need to use reinforcement learning (RL) over many millions of simulated interactions to build policies that can work across many situations. But this requires a very realistic physics simulation, otherwise the robots can learn behaviors that seem to work in simulation, but don’t apply in the real world. This is the sim-to-real gap. Agility uses advanced tools like NVIDIA’s Isaac Sim and Omniverse to simulate environments at a high-enough fidelity to avoid this gap: What works in simulation is also effective in the real world.
Tech Briefs: We are seeing impressive demos of humanoids trained in virtual worlds but reliability in messy real environments is what matters. What level of sim‑to‑real robustness is “good enough” for deployment, and how far away is that threshold today?
Velagapudi: In the demanding world of logistics and manufacturing, a 99 percent success rate is often viewed as a failure because even a 1percent error rate requires frequent human intervention, which undermines the return on investment. For meaningful commercial deployment, the industry requires multiple nines of reliability — a level where the robot quietly and consistently performs the same task thousands of times per shift without issue. While the current generation of AI and physical models are on the cusp of achieving human-level flexibility in reasoning, this technology is still very new and has not yet fully impacted commercial robotics at scale. Currently, Agility is one of the only companies to have moved beyond demos and pilots into real commercial deployment in production environments. However, the gap between impressive demonstrations and the near-perfect reliability required for widespread adoption remains a significant threshold that the industry is still working toward. Achieving this level of robustness is what separates a viral video from a viable business solution.
Tech Briefs: Digit is already being piloted in real warehouses, not just labs. How much of Digit’s behavior is still learned or validated in simulation, and how much now comes from direct experience on warehouse floors?
Velagapudi: Digit uses a combination of simulation, collected data, and engineered skills to perform tasks in real warehouses. Rather than a single large model, Digit is powered by a number of AI models that each focus on different layers of control. The right training approach varies by the layer of the stack. Simulation works best for low-level controls and motions, things like learning to walk, or carry objects. These basic motions need to be hardened through exploring the simulation space to figure out exactly the right ways to robustly execute motions. Above this sits the skills layer, which can be either learned from simulation, hand-engineered, or learned from datasets recorded from any Digit in the fleet. These datasets can come from test cells, deployed robots, or special training cells where humans control Digit directly. While we are teaching Digit new skills, much of its behavior is learned from data collected by robot test cells replicating customer sites located at each Agility office. This lets us carefully curate scenarios that we have seen in our years of collected robot logs from production deployments. Over time, this will increasingly shift toward training on the production data-lake, which includes logs from our years of robot deployments.
Tech Briefs: From a systems perspective, what metrics do you track internally to validate sim‑to‑real success for Digit?
Velagapudi: Internally, Agility focuses on metrics that define production-grade reliability and long-term utility in industrial settings. The three most important ones are uptime, performance, and reliability. In order to maximize them, we need to achieve repeatable material-handling performance and the ability to work multiple shifts consistently. We can then track the impact of our deployments with metrics such as shifts worked and number of items moved. Tracking the long tail of failure modes is critical; this involves analyzing how Digit recovers from real-world anomalies like broken totes or slippery floors. By monitoring these recoveries and the frequency of necessary interventions, Agility can iterate on both hardware and control systems to improve performance. Ultimately, the success of Digit is not measured by its ability to perform spectacular tricks, but by its ability to show up every day and perform the same monotonous tasks with the same predictability and efficiency as a traditional piece of industrial equipment.
Tech Briefs: In shared warehouse spaces, Digit has to balance efficiency with safety margins around humans. How do you design Digit’s movements, stopping behavior, and response to nearby workers so humans consistently feel safe around it?
Velagapudi: Although Digit, like all humanoids, is currently separated from humans in safety-guarded work cells, Agility’s goal is to achieve full collaborative safety, allowing robots to work side-by-side with people without barriers. In order to do this, we need the robot to be transparent and predictable. One of the most important design decisions we made to support this was treating Digit as industrial equipment from the start. That meant integrating familiar components used throughout manufacturing, such as emergency stop buttons, standard visual and audio indicators, and other common functional-safety semantics. These elements make Digit easier to integrate into existing factory safety architectures and allow operators to interact with the robot using tools and procedures they already understand. This means as Digit automates mind-numbingly repetitive tasks, it is viewed as a reliable and supportive machine that allows employees to focus on more complex, higher-value roles.
Tech Briefs: Looking ahead five years, what do you think will matter more for warehouse humanoids like Digit to scale — battery life, better actuator and hardware design, or safety?
Velagapudi: While advancements in simulation and hardware are important, safety is the most critical factor for scaling warehouse humanoids over the next five years. The real barrier is not AI or battery technology, but the ability to build a robot that can safely work side-by-side with humans without being cordoned off in restricted zones. Currently, the lack of dedicated safety standards for dynamic humanoid robots forces them to operate in separate work cells. Agility actively is working with standards bodies to develop the ISO/TC 299 and ISO 25785 standards, which will provide the necessary framework for cooperative safety (the first level of safety before full collaborative safety) that will allow the physical barriers separating humans from humanoid robots in a work environment to be removed. Achieving these certifications is the key to unlocking widespread market adoption, as it allows the robot to move freely through human-centric spaces. While improvements in battery life and payload capacity also are priorities, they are considered secondary to the challenge of building trust and ensuring safety. Scaling only will happen once humanoids can quietly become a safe, reliable part of daily operations.
Tech Briefs: Looking beyond warehouses, homes are often described as the ultimate challenge for humanoid robots. Based on real‑world deployment so far, what still has to change before humanoid robots can realistically work in homes?
Velagapudi: Transitioning from warehouses to homes presents a much greater challenge due to the unstructured and idiosyncratic nature of domestic environments. Homes are diverse, changing structures: We don’t think twice about moving furniture or leaving a pile of laundry out. They are also much more challenging to navigate safely — in a warehouse, you can guarantee that a robot is only interacting with working adults, while in a home you have children and pets who rarely act predictably. This leads to a conflict between productivity and safety. For a robot to be useful, it must be strong enough to lift a laundry hamper, yet gentle enough to avoid injuring children or pets if it falls. Building robots that we can trust in these environments will take time. This level of trust will only be reached after first seeing humanoid robots with near 100 percent reliability in industrial and commercial settings. While AI and reasoning capabilities are on the cusp of producing human-like flexibility, the complexity of performing diverse tasks in the home in a safe and reliable manner means that general-purpose home humanoids are likely many years away.
This article was written by Chitra Sethi, Editorial Director, SAE Media Group. For more information, visit here .
