The technical challenges and future of DePIN and embodied intelligence

! Technical Challenges and Future of DePIN and Embodied Intelligence

On February 27, Messari hosted a podcast on "Building Decentralized Physical AI" with Michael Cho, co-founder of FrodoBot Lab. They chatted about the challenges and opportunities of decentralized physical infrastructure networks (DePIN) in robotics. While still in its infancy, this field has great potential to revolutionize the way AI bots work in the real world. However, unlike traditional AI, which relies on large amounts of internet data, DePIN robotics AI technology faces more complex issues, such as data collection, hardware limitations, evaluation bottlenecks, and the sustainability of the economic model.

In today's article, we'll break down the key points in this discussion, look at the problems DePIN robotics encountered, what are the main barriers to scaling decentralized bots, and why DePIN is more advantageous than centralized methods. Finally, we'll explore the future of DePIN robotics to see if we're on the verge of a "ChatGPT moment" for DePIN robotics.

Where is the bottleneck of DePIN intelligent robots?

When Michael Cho first started working on FrodoBot, his biggest headache was the cost of robotics. The price of commercial robots on the market is ridiculously high, which makes it difficult to promote AI applications in the real world. His initial solution was to build a low-cost autonomous robot that would cost as little as $500, intending to win at a cheaper price than most existing projects.

But as he and his team worked deeper, Michael realized that cost wasn't really the bottleneck. The challenges of a decentralized physical infrastructure network (DePIN) in robotics are far more complex than "expensive or not". As FrodoBotLab continues to advance, multiple bottlenecks in DePIN robotics are emerging. To achieve large-scale deployment, the following bottlenecks must be overcome.

Bottleneck 1: Data

Unlike large 'online' AI models trained on large amounts of internet data, embodied AI (AI) requires interaction with the real world to develop intelligence. The problem is that there is no such large-scale basis in the world, and there is no consensus on how to collect this data. Data collection for embodied AI can be grouped into the following three broad categories:

▎The first category ishuman operation data, which is the data generated when humans manually control robots. This type of data is of high quality and captures video streams and motion tags—that is, what humans see and how they react accordingly. This is the most effective way to train AI to mimic human behavior, but it has the disadvantage of being costly and labor-intensive.

▎The second type is synthetic data (simulation data), which is very useful for training robots to move in complex terrain, such as training robots to walk on rough ground, which is very useful for some specialized fields. But for some of the most varied tasks, such as cooking, simulating an environment is not so good. We can imagine the situation of training a robot to fry eggs: small changes in the type of pan, oil temperature, room conditions can affect the results, and it is difficult for the virtual environment to cover all the scenes.

▎The third category is video learning, which is to let the AI model learn by observing real-world videos. While this approach has potential, it lacks the real physical direct interactive feedback required for intelligence.

Bottleneck 2: Level of autonomy

Michael mentions that when he first tested FrodoBot in the real world, it was mainly using robots for last-mile deliveries. Statistically, the results are actually pretty good – the robot successfully completed 90% of the delivery tasks. But a 10% failure rate in real life is unacceptable. A robot that fails every 10 deliveries is simply not commercial. Just like automated driving technology, autonomous driving can have a record of 10,000 successful drives, but one failure is enough to defeat the confidence of commercial consumers.

Therefore, for robotics to be truly useful, the success rate needs to be close to 99.99% or even higher. But the problem is that for every 0.001% improvement in accuracy, it takes exponential time and effort. Many people underestimate the difficulty of this final step.

Michael recalls that when he sat in Google's self-driving car prototype in 2015, he felt that fully autonomous driving was on the verge of becoming a reality. Ten years on, we're still debating when Level 5 will be fully autonomous. Advances in robotics are not linear, but exponential in nature – with each step forward, the difficulty increases dramatically. This last 1% accuracy rate can take years or even decades to achieve.

Bottleneck 3: Hardware: AI alone cannot solve the problem of robots

Taking a step back, even with the best AI models, existing robot hardware isn't ready for true autonomy. For example, the most overlooked problem in hardware is the lack of tactile sensors – the best current technologies, such as Meta AI's research, are nowhere near the sensitivity of a human fingertip. Humans interact with the world through sight and touch, while robots know little about texture, grip, and pressure feedback.

There's also the occlusion problem – when an object is partially blocked, it's hard for the robot to recognize and interact with it. And humans can intuitively understand an object even if they can't see it in its entirety.

In addition to the perception problem, the robot actuator itself is also flawed. Most humanoid robots place their actuators directly on their joints, making them bulky and potentially dangerous. In contrast, the human tendon structure allows for smoother and safer movements. That's why existing humanoid robots look stiff and inflexible. Companies like Apptronik are developing more bio-inspired actuator designs, but these innovations will take time to mature.

Bottleneck 4: Why is hardware expansion so difficult?

Unlike traditional AI models, which rely solely on computing power, the implementation of intelligent robotics requires the deployment of physical devices in the real world. This poses a significant capital challenge. Building robots is expensive, and only the richest big companies can afford large-scale experiments. Even the most efficient humanoid robots now cost tens of thousands of dollars, making mass adoption simply unrealistic.

Bottleneck 5: Evaluate effectiveness

This is an "invisible" bottleneck. If you think about it, a large online AI model like ChatGPT can test its functionality almost instantaneously – after a new language model is released, researchers or ordinary users around the world can draw conclusions about its performance in a matter of hours. But evaluating physical AI requires real-world deployments, which take time.

Tesla's Full Self-Driving (FSD) software is a good example. If Tesla records 1 million miles with no accidents, does that mean it has really reached Level 5 autonomy? What about 10 million miles? The problem with robotic intelligence is that the only way to validate it is to see where it ultimately fails, which means large-scale, long-term, real-time deployments.

Bottleneck 6: Manpower

Another underestimated challenge is that human labor remains indispensable in robotic AI development. AI alone isn't enough. Robots need training data from human operators; The maintenance team keeps the robot running; and essential researchers/developers to continuously optimize AI models. Unlike AI models that can be trained in the cloud, bots require constant human intervention – a major challenge that DePIN must address.

The Future: When Will the ChatGPT Moment for Robotics Arrive?

Some believe that the ChatGPT moment for robotics is coming. Michael is somewhat skeptical. Given the hardware, data, and evaluation challenges, he believes that general-purpose robotic AI is still far from mass adoption. However, the progress of DePIN robotics does give some hope. The development of robotics should be decentralized and not controlled by a few large companies. The scale and coordination of a decentralized network can spread the burden of capital. Instead of relying on a large company to pay for thousands of robots, put individuals who can contribute into a shared network.

To illustrate – first and foremost, DePIN accelerates data collection and evaluation. Instead of waiting for a company to deploy a limited number of bots to collect data, decentralized networks can run in parallel and collect data on a much larger scale. For example, in a recent AI-to-human robotics competition in Abu Dhabi, researchers from institutions such as DeepMind and UT Austin put their AI models to the test against human players. While humans still prevail, the researchers are excited about the unique datasets collected from real-world robot interactions. This is a testament to the need for subnets that connect the various components of robotics. The enthusiasm of the research community also shows that even if full autonomy remains a long-term goal, DePIN robotics has demonstrated tangible value from data collection and training to real-world deployment and validation.

On the other hand, AI-driven hardware design improvements, such as optimizing chips and materials engineering with AI, could significantly shorten the timeline. A concrete example is when FrodoBot Lab partnered with other institutions to secure two boxes of NVIDIA H100 GPUs—each containing eight H100 chips. This provides researchers with the necessary computing power to process and optimize AI models for real-world data collected from robot deployments. Without such computing resources, even the most valuable datasets cannot be fully utilized. With access to DePIN's decentralized computing infrastructure, the robotics network allows researchers across the globe to train and evaluate models without being constrained by capital-intensive GPU ownership. If DePIN succeeds in crowdsourcing data and hardware advancements, the future of robotics could come sooner than expected.

In addition, AI agents like Sam (a traveling KOL bot with meme coins) demonstrate a new monetization model for decentralized robotics networks. Sam operates autonomously, streaming live 24/7 in multiple cities, and its meme coins are also increasing in value. This model shows how intelligent bots powered by DEPIN can sustain their finances through decentralized ownership and token incentives. In the future, these AI agents could even use tokens to pay for assistance from human operators, rent additional bot assets, or bid on real-world tasks, creating an economic cycle that benefits both AI development and DePIN participants.

Summary

The development of robot AI depends not only on algorithms, but also on hardware upgrades, data accumulation, financial support, and human involvement. In the past, the growth of the robotics industry was limited by high costs and the dominance of large enterprises, which hindered the speed of innovation. The establishment of the DePIN bot network means that with the power of the decentralized network, robot data collection, computing resources, and capital investment can be coordinated on a global scale, not only accelerating AI training and hardware optimization, but also lowering the development barrier to allow more researchers, entrepreneurs, and individual users to participate. We also expect that the robotics industry will no longer rely on a few tech giants, but will be driven by the global community to move towards a truly open and sustainable technology ecosystem.

*All content on the Coinspire platform is for informational purposes only and does not constitute an offer or recommendation of any investment strategy, and any individual decisions made based on the content of this article are the sole responsibility of the investor, and Coinspire is not responsible for any gains or losses arising therefrom.

Investment is risky, and decisions need to be made carefully