Figure AI's Humanoid Robot Achieves Natural Walking Through AI Simulation

Fancy humanoid robot no longer walks like it urgently needs a toilet

'Years' of training in simulation helped the Figure bipedal robot walk more like a real human. Human-looking bipedal robots can already run, jump, breakdance, punch, and generally perform broad feats of athletic prowess most humans could only dream of. One thing they are still pretty bad at though is walking a straight line without looking like they are moments away from soiling themselves. Figure AI, one of the buzziest startups in the humanoid robot space, now says it has engineered a solution to help address their machine's stiff shuffle-step. The more natural-looking stride was achieved by analyzing thousands of virtual humanoid robots walking simultaneously in a simulated digital environment, Figure explained in a recent blog post. The company used reinforcement learning, rewarding the virtual robots for actions like synchronized arm swings, heel strikes, and toe-offs (when the toe leaves the ground) that more closely resemble human movement. Figure says this simulated approach allows it to generate years worth of data in just days. The refined "Learned Natural Walking" model was then applied to a physical Figure 2.0 humanoid. As the demonstration video above shows, the new stride isn't perfect, but it's an improvement over what the company's robots displayed just a month earlier. The latest version moves more naturally, especially in its heel and lower leg movements. Figure exposed its horde of simulated robots to a variety of different terrains. It also had them adapt to unexpected scenarios like being tripped or shoved. The company sees this all as a step toward making its robots move more like humans in a real, unpredictable physical world. "These initial results are exciting, but we believe they only hint at the full potential of our technology," Figure writes. "We're committed to extending our learned policy to handle every human-like scenario the robot might face in the real world." Research suggests that humans tend to respond more positively to robots when they appear more human-like and can hold a conversation. This same trait also makes people less likely to bludgeon the machines -- a factor that could become increasingly important as humanoid robot manufacturers push for wider adoption across various industries. Figure AI has already secured a deal with BMW to test its robots in a South Carolina manufacturing facility, while Amazon is reportedly trialing humanoid robots from Agility Robotics in some of its warehouses. Humanoid robots have been performing backflips and dance routines for years, but they often struggle with the simple tasks humans take for granted. This contrast -- where robots excel at challenges humans find difficult but fail at tasks humans find easy -- is often referred to in robotics as "Moravec's Paradox." While advances in AI models have helped narrow this gap in recent years, there's still a long way to go. Google DeepMind just last year made a major breakthrough by finally teaching a robot how to tie a shoe. Even then, after decades of robotics research, the robot did a worse job than some school-age children. At the same time, robots today still struggle to lightly grip various objects without crushing them and often have about as much sense of balance as a human after one too many margaritas.

Watch eerie video of humanoid robot 'army' marching naturally, thanks to a major AI upgrade

Figure 02's human-like gait is the product of the company's simulated reinforcement learning system, and is just the beginning of its plans to make its robots perform physical tasks more naturally. A U.S. robotics company has used artificial intelligence (AI) to give its humanoid robots a more natural-looking stride, and they say it's just the beginning. In the promotional video, the robot, called Figure 02 and manufactured by the company Figure, marches with a "human-like" gait. This is an ability it claims will help its robot to navigate the physical world more easily. "These initial results are exciting, but we believe they only hint at the full potential of our technology," company representatives wrote in a blog post accompanying its announcement. "We're committed to extending our learned policy to handle every human-like scenario the robot might face in the real world." Robots have been running, cartwheeling, doing backflips, breakdancing, and beating us at chess for years now. But performing tasks that appear simple to humans, such as walking in a straight line, gripping objects, tying shoelaces, and navigating social situations, have proven tough for robots to master. Related: Chinese humanoid robot is the 'fastest in the world' thanks to its trusty pair of sneakers The problem, known as Moravec's Paradox, emerges because computers excel at problems that require complex calculations and large datasets, but lack our real-world experience honed by millions of years of evolution. This makes robots' shuffling gaits, well, robotic at best. At worst, it gives them the appearance that they may have soiled themselves. To tackle the robot's unnatural gait, Figure's engineers used a learning technique called reinforcement learning -- placing thousands of virtual robots inside a physics simulator that recreates various terrains, thereby improving their walking through trial and error. By rewarding the virtual robot army for natural motions, they refined their gaits to appear more human-like. With this task accomplished, they uploaded the refined "Learned Natural Walking" model to a real-world Figure 02 robot. The result is an android that can move somewhat naturally, with heel strikes, toe-offs and synchronized arm swings. Figure's reinforcement learning technique is key to the California company's plans to roll out its robots on factory floors. It has already tested its humanoid robots in a BMW factory in 2024 and plans to introduce more this year. Meanwhile, Apptronik, a Texas-based competitor, is also commercializing its humanoid robot, Apollo, for use in Mercedes-Benz factories by the end of 2025. Agility Robotics' Digit will also be introduced into warehouses this year.

Figure's humanoid robot no longer walks like it needs the bathroom

We've recently seen humanoid robots that can cartwheel, kung-fu kick, and front flip, but such attention-grabbing stunts aren't the goal of California-based Figure AI. Instead, its team of roboticists is focusing on designing an AI-powered bot that can move quickly and reliably and get things done. In a video (top) shared on X on Tuesday, Figure showed its own humanoid 02 robot performing "learned, natural walking." Figure's footage demonstrates the scale of the improvement in walking ability achieved by the team behind the robot. As you can see, the original 01 robot had more of a waddle about it, similar to how you might move if you were desperate for the bathroom. The latest 02 design, however, has a more relaxed walking style, with more realistic strides that help it to move more quickly -- important for when the bipedal bot is deployed in the workplace or the home. Indeed, Figure said in a post on X on Tuesday that this year is set to be "a big one" as it's "launching into production manufacturing, scaling up robots at our commercial customers, and working on launching robots into the home." A few weeks ago, Figure CEO Brett Adcock revealed that Helix -- the AI model that Figure uses to power its humanoid robot -- was advancing more quickly than expected, enabling the team to accelerate its timeline for home deployment by two years, meaning that testing will begin sometime this year. Figure's impressive 02 robot stands at 5 foot and 6 inches (168 centimeters), tips the scales at 154 pounds (70 kilograms), and can function for about five hours on a single charge. The company has already completed a trial deployment of its humanoid robot at a BMW facility in South Carolina in which a number of its robots were used to place sheet metal parts into specific fixtures that were then assembled as part of a vehicle's chassis. Figure says its overall ambition is "to develop general purpose humanoids that make a positive impact on humanity and create a better life for future generations," adding that its AI-powered designs "can eliminate the need for unsafe and undesirable jobs -- ultimately allowing us to live happier, more purposeful lives."

Figure AI's 02 Humanoid Can Now Walk Naturally Like a Human

The company has achieved this through an end-to-end neural network trained using reinforcement learning. Figure AI, the California-based robotics company, has introduced a humanoid robot capable of walking with human-like movements. The breakthrough, announced on Tuesday, aims to enhance the robot's adaptability for industrial and domestic applications by mimicking natural human locomotion. The innovative approach compresses years of simulated training into hours. Using reinforcement learning (RL), the robot learns through trial-and-error simulations to balance, shift weight, and walk naturally. CEO Brett Adcock took to X to announce this, saying, "Say goodbye to the Biden Walk!" This training is transferred directly to real-world hardware using domain randomisation and high-frequency torque feedback, enabling zero-shot deployment without additional fine-tuning. Figure AI's training process simulates thousands of virtual humanoids in parallel under varied physical parameters and scenarios. These include changes in terrain, actuator dynamics, and external disturbances like trips or slips. The Figure 02 robot achieves this through an end-to-end neural network trained via reinforcement learning (RL) in a high-fidelity physics simulator. As per the company's announcement, the robots are exposed to a wide range of scenarios they might encounter, and a single neural network policy learns to operate them all. The robots now exhibit features such as heel strikes, toe-offs, and synchronised arm swings. According to the company, these improvements were achieved by rewarding the robot for mimicking human walking trajectories while optimising for velocity tracking, energy efficiency, and robustness. Figure AI showcased 10 Figure 02 robots operating on the same RL neural network without modifications. This consistency emphasises scalability across its fleet of robots without manual adjustments. "This gives us hope this process can scale to thousands of Figure robots in the near future," the company stated. Adcock also added that "2025 will be a pivotal year as we start production, ship more robots, and tackle home robotics." The company mentioned in a post on X, "This year is going to be a big one for Figure. We're launching into production manufacturing, scaling up robots at our commercial customers, and working on launching robots into the home." The technology addresses labor shortages and safety concerns while opening possibilities for broader applications. The company plans to scale production, manufacturing, and commercial deployment later this year. This development positions Figure AI as a competitor in the humanoid robotics space alongside Tesla's Optimus and Agility Robotics' Digit. Not just these, but also alongside Chinese competitors like UBTECH Robotics and Unitree Robotics. The company recently announced Helix, a Vision-Language-Action (VLM) model that allows humanoid robots to perform complex tasks using natural language. This marked a leap in AI for humanoids before any other company. This model marked progress in robotics, enabling robots to understand and react to instructions in real time, handle unforeseen objects, and collaborate.

Figure AI Uses Reinforcement Learning to Teach Robots to Walk Naturally

Figure AI has developed a new humanoid robotic natural walking capability for its humanoid robots, leveraging reinforcement learning (RL) and simulation-based training. This approach enables the humanoid robots to walk with human-like gait patterns, ensuring robust and adaptable locomotion across various environments. The technology allows for efficient scaling and deployment across a fleet of robots without additional tuning. Reinforcement learning serves as the foundation of Figure AI's approach to humanoid locomotion. This machine learning technique enables robots to learn through trial and error, guided by reward signals that prioritize key objectives such as stability, energy efficiency, and speed. In virtual environments, thousands of simulated humanoids undergo rigorous testing to identify optimal walking strategies. These strategies are refined through iterative processes, allowing robots to adapt to complex scenarios, including uneven terrain and sudden disturbances. For instance, virtual robots are trained to recover from slips or trips, making sure resilience in unpredictable conditions. By using reinforcement learning, Figure AI minimizes the need for extensive physical testing, significantly accelerating development timelines while conserving resources. This method not only enhances the robots' adaptability but also ensures that they are well-prepared for real-world challenges. High-fidelity simulations play a pivotal role in the natural walking robots training process for Figure AI's robots. These virtual environments replicate real-world conditions with exceptional accuracy, allowing robots to gain years of walking experience within a matter of hours. Through simulation, robots are exposed to a wide range of terrains, actuator dynamics, and external forces. This comprehensive training ensures that their learned behaviors are both versatile and reliable. For example, robots are trained to navigate surfaces ranging from smooth indoor floors to rugged outdoor paths, preparing them for deployment in environments such as warehouses, construction sites, and urban streets. Simulation-based training offers several advantages: This approach underscores the importance of virtual environments in advancing humanoid robotics, providing a controlled yet dynamic platform for experimentation and refinement. One of the most notable features of Figure AI's robots is their ability to replicate human-like walking patterns. These robots exhibit biomechanically accurate movements, including heel strikes, toe-offs, and synchronized arm-leg coordination. Such naturalistic movements enhance their functionality, making them more suitable for human-centric environments. The training process involves balancing multiple objectives to achieve optimal performance. For example, robots are programmed to minimize energy consumption while maintaining smooth and stable transitions between steps. This dual focus on efficiency and stability not only improves their operational capabilities but also reduces mechanical wear, extending their lifespan. By achieving a human-like gait, Figure AI's robots demonstrate a level of sophistication that enhances their usability in tasks requiring precision and adaptability. A critical challenge in robotics is making sure that behaviors learned in simulations translate effectively to physical robots. Figure AI addresses this challenge through domain randomization, a technique that introduces variability into simulations to account for real-world uncertainties. By training robots under diverse conditions, the system ensures that their learned behaviors are robust and adaptable. Additionally, the robots use high-frequency torque feedback to adjust their movements in real time. This feedback mechanism compensates for discrepancies between simulated and physical environments, such as variations in friction or actuator dynamics. The result is a seamless transition from virtual training to physical deployment, minimizing the need for manual adjustments and making sure consistent performance across different settings. Figure AI's technologies are designed with scalability in mind, allowing the deployment of a fleet of humanoid robots that maintain consistent performance. The reinforcement learning-driven approach standardizes training, allowing multiple robots to operate autonomously without requiring extensive customization. This scalability is particularly advantageous for commercial applications. For instance, a fleet of humanoid robots could efficiently manage inventory in a warehouse, with each robot performing tasks independently while adhering to uniform performance standards. Key benefits of scalable deployment include: By streamlining deployment, Figure AI positions itself as a leader in the commercialization of humanoid robotics. Figure AI's vision centers on creating versatile natural walking robots capable of performing human-like tasks with natural movement. This vision is supported by a commitment to rapid iteration, real-world deployment, and the integration of advanced technologies. The company's approach combines reinforcement learning, simulation-based training, and scalable systems to push the boundaries of humanoid robotics. These advancements not only enhance the capabilities of humanoid robots but also pave the way for their widespread adoption in commercial and industrial settings. Potential applications for these robots span a wide range of industries, including: As Figure AI continues to refine its technologies, the potential for humanoid robotics to transform industries becomes increasingly evident. By addressing key challenges and using innovative solutions, the company is shaping the future of robotics with a focus on practicality and adaptability. Here are additional guides from our expansive article library that you may find useful on humanoid robots :