DePAI: Crypto's Role in Scaling Humanoid Robotics

Humanoids: From Fiction to Frontline

Commercialized robotics is not a new concept. Most are familiar with products such as the iRobot Roomba vacuum, debuting in 2002, or more recent household robotics such as Kasa’s pet camera. Both are single-purpose built. With the help of AI, robots are evolving from single-purpose machines to multi-purpose, designed to operate in unstructured environments.

Humanoid robots will progress from basic tasks like cleaning and cooking to concierge, firefighting, and even surgery over the next 5–15 years.

Recent developments are turning humanoid robotics from science fiction to reality.

• Market expansion: 100+ companies building humanoids (e.g. Tesla, Unitree, Figure, Clone, Agile, etc.).
• Hardware tech has crossed the uncanny valley: New humanoids exhibit fluid, natural motion, enabling human-like interactions in real-world environments. Unitree’s H1 walks faster than the average human (3.3 m/s vs 1.4 m/s).
• New paradigm of labor costs: Humanoid robots are expected to undercut U.S. wages by 2032.

The Bottleneck: Real-World Training Data

Despite clear tailwinds for humanoid robotics, mass deployment remains bottlenecked by data quality and scarcity.

Other AI embodiments, like autonomous driving, have largely overcome the data issue via cameras and sensors on existing vehicles. In the case of self-driving (e.g., Tesla, Waymo), these fleets are able to generate billions of miles of real‑world driving data. Waymo was able to put their cars on the road for real-time training with a human “babysitter” in the passenger seat during this phase.

However, consumers are unlikely to tolerate the presence of a “robot babysitter”. Robots must be performant out-of-the-box, making pre-deployment data acquisition essential. Training must be completed prior to commercial production, where scale and quality of data persist as an issue.

While each training modality has its own unit of scale (i.e. tokens for LLMs, video-text pairs for image generators, and motion episodes for robotics), the below comparison highlights the orders-of-magnitude gap in data availability that robotics data is contending:

• GPT-4 was trained on over 15 trillion text tokens
• Midjourney/Sora leverage billions of labeled video-text pairs
• In contrast, the largest robotics datasets contain only ~2.4 million episodes

This disparity illustrates why robotics has yet to achieve a true foundation model in the same way LLMs have. The data simply isn’t there yet.

Traditional data collection methods do not scale for humanoid robotics training data. Current methods include:

• Simulation: Cheap but lacks real-world edge cases (Sim2Real gap)
• Internet Video: Doesn’t contain proprioceptive or force feedback context needed for robot learning
• Real-world data: Accurate, but requires teleoperation and a human in the loop, which is costly ($40K+ per robot) and lacks scalability.

Training in virtual environments is inexpensive and scalable, but models often struggle when deployed in the real world. This problem is known as the Sim2Real gap.

For example, a robot trained in simulation might succeed at picking up objects with perfect lighting and flat surfaces, but fail when faced with cluttered environments, uneven textures, or imperfect situations humans are accustomed to in the physical world.

Reborn provides a way to cheaply and quickly crowdsource real-world data, enabling robust robotics trainingand solving the Sim2Real gap.

Reborn: A Full-Stack Vision for Decentralized Physical AI

Reborn is building a vertically integrated software and data platform for physical AI. At its core, Reborn is solving the data bottleneck for humanoid robotics, but its ambition extends far beyond that. Through a combination of proprietary hardware, multi-modal simulation infrastructure, and foundation model development, Reborn becomes a full-stack enabler of embodied intelligence.

The Reborn stack begins with the “ReboCap”, a proprietary consumer-grade motion capture device. This powers a rapidly growing AR/VR gaming ecosystem, where users generate high-fidelity motion data in exchange for network incentives. Reborn has sold over 5,000 ReboCap units and now supports 160,000 monthly active users (MAUs), with a clear path toward two million by year-end.

Impressively, this growth has been organic: users are drawn to the entertainment value of the games themselves, and livestreamers are adopting ReboCap to animate digital avatars with real-time body tracking. This organic engagement loop powers scalable, low-cost, and high-fidelity data generation, making Reborn’s dataset a valuable training resource for leading robotics companies.

The second layer of Reborn’s software stack is Roboverse, a multi-modal data platform that unifies fragmented simulation environments. Today’s simulation landscape is highly fragmented, for example tools like Mujoco and NVIDIA Isaac Lab each offer different strengths but lack interoperability. This balkanization slows progress and exacerbates the Sim2Real gap. Roboverse addresses this by standardizing across simulators, creating a shared virtual infrastructure for developing and evaluating robotics models. This integration allows for consistent benchmarking, improving scalability and generalizability.

Together, ReboCap and Roboverse form the base of Reborn’s full-stack platform. The former captures real-world data at scale, while the latter orchestrates simulation environments for model training. This integrated approach showcases the true power of Reborn’s DePAI network. It is building a developer platform for Physical AI that extends beyond simple data acquisition, to actual model deployment and licensing.

The Reborn Foundation Model

Perhaps the most crucial component of Reborn’s software stack is the Reborn foundation model (RFM). Reborn is building one of the first robotics foundation models, designed to serve as core infrastructure for the emerging Physical AI stack. Think traditional foundation models for LLMs, such as OpenAI’s o4 or Meta’s Llama, but for robots.

The combination of the three major elements of Reborn’s stack (ReboCap, Roboverse, and the RFM), creates a strong vertically integrated moat for Reborn. By pairing crowdsourced motion data with robust simulation and model licensing, Reborn can train models with the scale and diversity needed to generalize across use cases. The result is a foundation model that supports downstream applications in a wide-range of use-cases, including industrial, consumer, and research robotics.

Reborn is actively commercializing its technology, launching paid pilots with Galbot and Noematrix and establishing strategic partnerships with Unitree, Booster Robotics, Swiss Mile, and Agile Robots. China's humanoid robot market is experiencing rapid growth, accounting for around32.7% of the global market. Notably, Unitree holds over 60% of the global quadruped robot market and is among six Chinese humanoid robot manufacturers planning to produce over 1,000 units in 2025.

Crypto's Role in the Physical AI Stack

Crypto is enabling the full vertical stack for physical AI.

While all these projects sit in different parts of the physical AI stack, they all share something in common: 100% of them are DePAI projects! DePAI makes decentralized physical AI possible by ensuring open, composable, and permissionless scaling via token incentives across the stack.

The fact that Reborn hasn’t launched a token yet makes its organic growth even more impressive. Once token incentives go live, network participation is expected to accelerate as a part of the DePAI flywheel: Reborn issues incentives for acquiring its hardware (the ReboCap), robotics companies pay ReboCap owners for their contributions, encouraging more folks to purchase and use ReboCap. Reborn will also dynamically incentivize high-value edge case behaviors – ensuring even better coverage of the Sim2Real gap.

It's all about the Data

Robotics “ChatGPT” moment won’t come from robotics companies themselves because hardware is much tricker to deploy than software. Virality in robotics is inherently constrained by cost, hardware availability, and logistical complexities. These factors are absent in purely digital software like ChatGPT.The tipping point for humanoid robotics will come not when prototypes impress, but when costs fall enough for mass adoption — as with smartphones or PCs. When costs fall, hardware becomes tablestakes. The real competitive edge will lie in data and models. Specifically, the scale, quality, and diversity of motion intelligence used to train these machines.

Conclusion

The robotics platform shift is inevitable but, like all platforms, it needs data to scale. Reborn is a high-leverage bet that crypto can fill the most acute gap in the AI robotics stack. DePAI for robotics data is cost-efficient, scalable, and composable. In a world where robotics are the next frontier of AI, Reborn is the equivalent of turning everyday humans into the “miners” of motion data. As LLMs need text tokens, humanoid robots need motion episodes. Reborn is how we unlock one of the last remaining bottlenecks in turning humanoid robotics from sci-fi to reality.

Author: @brezshares with input from @CabronElBufon @jackbplatts and @MehdiFarooq2

Special thanks to the @reborn_agi team for their review and insight!