Blockchain, agents, and robots: from the virtual world to the real one

Broadly speaking, AI devices refer to hardware — including robots, vehicles, and even household appliances related to the IoT — that is run by artificial intelligence algorithms.

Like virtual agents, their physical “siblings” require mature infrastructure and complex technical solutions. In response to this demand, a dedicated segment has emerged in the crypto industry known as DePAI.

Its goal is to offer a decentralized alternative to proprietary AI-device solutions, whether that means compute power or training data. 

Developers of such projects argue that blockchain can provide AI devices with the following advantages:

• no single point of failure — a critical factor when it comes to managing complex machine networks or storing user data; 
• transparency — all device actions are recorded and available for verification. An important aspect of security or validating results;
• coordination infrastructure — data exchange, payments, and other interactions can be carried out on-chain and automated via smart contracts;
• economic incentives — tokenization and rewarded activities make expensive hardware solutions more accessible, and can also speed up training data collection or incentivize product adoption.

The DePAI concept is closely tied to DePIN, which focuses on building distributed technical infrastructure for everyday use — cellular networks, geolocation systems, mapping, and so on. 

That is why a number of projects started out as purely infrastructure plays, but found new use cases as artificial intelligence evolved. 

For example, Akash Network was conceived as a decentralized platform for GPU computing, but because this hardware is well-suited to serving large language models (LLMs), the project is often categorized as DePAI. Similarly, Hivemapper was built for mapping, but since this data is used for AI training or navigation, it is also included in the new category.

At the same time, general-purpose platforms are being complemented by specialized ones aimed at solving specific problems that arise in machine development and interaction.

DePAI infrastructure includes several components required for a “thinking” machine to operate. Together, they form a hierarchical system, where the lower layers contain data collection and compute solutions, and the upper layers contain coordination protocols and decision-making modules.

The ultimate goal of this ecosystem is to create what is known as robonomics. It is an analogue of the agent economy discussed in the first part of the article, but its participants are not programs — they are machines: self-driving cars, drones, industrial and humanoid robots, and, in a broader sense, all IoT devices.

Let’s break down the key infrastructure segments and the projects operating within them.

Training and data

In theory, control of a device can be delegated to a virtual agent. It would be responsible for decision-making — for example, whether a machine should provide a certain service or purchase a product — as well as for personalizing communication. 

A similar product was planned on the basis of ElizaOS, but it never progressed beyond the announcement. The problem is that these programs “don’t know” how to control hardware, which is why robot manufacturers like Tesla or Boston Dynamics use specialized, closed AI models. 

Training such algorithms requires a massive dataset about the physical world — from coordinates to images of objects. And DePAI provides efficient solutions for collecting it. 

For example, the Geodnet project is building a decentralized high-precision positioning network. It already includes more than 200 coordinate reference stations, whose operators receive rewards. This data is then used by devices to determine their location and navigate.

Another example is Auki. The platform enables agents to share visual data from their cameras to create a virtual map of the environment. This is a more complex development that serves both as a data layer and as a platform for robot coordination. 

Similar goals are pursued by NATIX Network, the previously mentioned Hivemapper, and Over The Reality. The data aggregated by these projects can be used to train AI in spatial perception and physical object recognition. 

Another interesting project is Silencio. This platform also operates through a distributed network of independent nodes, but instead of visual data, it collects sounds. The developers claim that this layer of information about the outside world is no less important for machines than imagery, especially in urban environments.

The key advantage of DePAI in all these cases is providing economic incentives for data collection and sharing. Blockchain enables ordinary people to record audio or video and get rewarded for their actions, whereas centralized companies would have to build their own node network with the associated costs.

Computing and information storage

We have already covered the problem of data storage for virtual agents. DePAI faces the same constraints and can use the same solutions, including specialized blockchains and data availability layers like EigenDA. With the caveat that there may be an order of magnitude more information. 

Computing is more challenging, because machines require far more power than software and are highly dependent on response times. As solutions, you can use DePIN projects, such as Akash Network, but there are also specialized platforms.

For example, Edge Network is building a distributed compute network to process requests from IoT devices and robots. The idea is to leverage nodes (users’ gadgets) in different locations and place them as close to the machines as possible, thereby reducing transmission latency. This should both remove the limitations of local computing and offset the drawbacks of remote cloud infrastructure. 

GAIB takes a different approach. The platform enables participants to co-own GPU farms through tokenization. These resources are used for AI-related computations, and the revenue from providing them is distributed among token holders. This model can incentivize the development of alternative base infrastructure that is not limited to data centers.

Worth highlighting separately is CodecFlow — a platform for task orchestration with automatic distribution of compute between local and cloud resources. It can also be used to build and train AI agents, including physical ones.

Identification and interaction

One of the most prominent projects in this area is FABRIC. It is a decentralized network by OpenMind, where each participant (robot) is identified using an ERC-7777 standard token. The latter contains a unique, verifiable digital signature generated on the device and can also include specific rules for behavior or interaction with humans. 

Connected machines can share data about their location and current task to coordinate actions.

OpenMind has also developed the OM1 information exchange protocol. It enables devices from different manufacturers to standardize the data transfer format. In addition, the company’s website lets you buy robots compatible with the standard and find instructions for installing OM1. 

Another example is RoboStack. It is a cloud platform for simulating a machine’s operation in a virtual environment for development and pre-deployment testing. One of the project’s solutions is RCP (Robot Context Protocol) — a standardized communication layer for robots, AI agents, and people. 

More notably, however, the protocol is built on Virtuals’ technology, meaning it is already integrated into the decentralized AI ecosystem and enables the platform’s agents to be used to “personalize” compatible devices. 

As a niche solution, it is also worth highlighting the Peaq blockchain, designed to provide payment infrastructure for AI devices. That said, this area is much better developed in the virtual agents sector, and the question of whether physical devices need a separate financial layer remains open.

There are also “universal” platforms — SingularityNET and Fetch.ai. We already mentioned them in the first part of the article, but they are partially related to DePAI. The first project works closely with Hanson Robotics, while the second has entered into a partnership with Bosch to develop new business models using IoT devices. 

It is important to understand that none of these solutions are integrated into a cohesive framework. Even if, taken together, they make it possible to launch or train a model optimized for DePAI, doing so would require bringing in separate services and manually “stitching” them together. Only OpenMind offers a more or less end-to-end approach, but in terms of tooling and usability, it is still far behind Eliza or Virtuals.

As of writing, the most viable DePAI projects are platforms that provide data for training AI devices. In this area, distributed infrastructure genuinely offers working solutions and products. But when it comes to controlling robots or coordinating them, blockchain platforms run into a number of obstacles.

Closed Infrastructure

The lion’s share of robotics is concentrated among centralized tech giants. Their developments and solutions — from motion-control models to spatial navigation systems — are typically proprietary. This also applies to robotaxis and even many “smart” home systems that control household appliances. 

Open solutions from Unitree or Boston Dynamics really do make it possible to plug a third-party “brain” into robots. However, the results of this approach are mixed. 

DePAI platforms can collect datasets for training or information about the environment, but all of this is used by third-party companies, because blockchain projects are not yet capable of building their own devices. That is why low-level infrastructure is developing faster than high-level infrastructure.

Technical Constraints

To move even individual aspects of AI device operations — such as computation or activity logging — onto a blockchain, you need scalable, low-cost networks, as well as large databases and compute hardware. 

In the first part of the article, we noted that hybrid solutions are often used for virtual agents, with data stored off-chain, and that specialized blockchains are only just starting to emerge. At the same time, activity in the “real” world generates even more data and requires corresponding throughput. 

Developers already have a number of proposals for coordinating the actions of physical agents using blockchain, including aerial drones and industrial robots, but so far only Fetch.ai has demonstrated a practical use case, and even then at a relatively small scale.

In addition, when controlling machines, request processing and response speed play a critical role. When it comes to a car, even milliseconds can determine the outcome of a maneuver. Whether modern blockchains can handle real-time transactions for thousands of devices remains an open question.

Security Requirements

Direct interaction between AI devices and people or objects increases the potential danger in the event of a targeted attack or an operational failure, due to the possibility of causing physical harm. At the same time, smart contract-based solutions are still vulnerable to hacks, and decentralized infrastructure is potentially susceptible to manipulation, for example, through the concentration of governance tokens or node compromise.

This aspect is also tied to regulation, because if something goes wrong, there must be an entity responsible for the consequences, whereas many blockchain projects do not have their own legal entity.

Regulatory Challenges

AI-based products do not have a dedicated legal status in most jurisdictions. Even the development of LLMs and interaction with them is often governed by general rules covering digital content or internet infrastructure. 

The problem only gets more complicated when it comes to physical devices — at the time of writing, only a few major cities like Los Angeles, Abu Dhabi, and Shanghai have allowed autonomous transportation. Many others, including San Francisco, Dubai, New York, and Washington, have introduced restrictions or an outright ban on such services. 

It is still hard to imagine how the use of “general-purpose” robots will be regulated. 

On top of that, there is regulatory uncertainty around decentralized platforms, open-source products, and tokenized digital assets. In the US, a bill intended to close the resulting gaps is still awaiting Senate approval.

A simple example: a Unitree robot running an open LLM like DeepSeek, owned by a DAO, damages a car. Who is liable — the manufacturer, the OS developer, or the owners? Only judges and lawmakers will be able to answer that question. 

It is also important to understand that, unlike the virtual agent sector, in physical AI, tech giants have a key advantage — a manufacturing monopoly. Only the developers decide where their devices’ data will be stored, what software is compatible with them, and whether third-party solutions can be connected at all.

On the other hand, centralized infrastructure, like DePAI, is still at an early stage of development. If blockchain’s advantages outweigh its constraints, and regulation allows distributed solutions, we may see a full-fledged synergy between the crypto industry and physical AI. 

But, as noted at the beginning of the article, only a “ChatGPT moment” can provide the spark to accelerate development. And we are still waiting for it to arrive.