Robots On-Chain: Crypto Powering the Physical-AI Era
VeradiVerdict - Issue #333
Summary
VLA innovations and economies of scale have catalysed the creation of affordable, efficient and generalised humanoid robots.
Robotics safety, financing and evaluation are worth exploring as warehouses expand into consumer robotics.
Crypto will improve robotics by providing economic guarantees for robot safety, and optimising its docking infrastructure, latency and data collection pipelines.
ChatGPT rewrote humanities’ expectations for AI. With the meta of LLMs interacting with the external software world, a lot of people thought AI Agents are the endgame. But when you look at iconic sci fi movies such as Star Wars, Blade Runner or Robocop, it is clear that humanity dreams of the day AI can physically interact with the world – in the form of robotics.
At Pantera Capital, we believe the ChatGPT moment for robotics is on the horizon. First, we will examine how the affordability and advancements in AI changed the space over the past few years. Then we would discuss how battery, latency and data collection optimization will change the space in the coming years and the role crypto plays in it. We would end by explaining why we believe robotics safety, financing, evaluation and education are verticals worth looking at.
What Changed?
AI Breakthroughs
Developments in the multi-modal LLM space is giving robots the brain it needs to carry out complex tasks. Robots mostly perceive its environment through two senses – visual and auditory.
Traditionally, visual models such as convolutional neural networks were tailored for object detection or classification tasks, but they lack the understanding to convert such visions into purposeful actions. LLMs are great at text-based understanding or generation, but they are restricted in their perception of the physical world.
Source: https://arxiv.org/html/2505.04769v1
Using Vision-Language-Action models (VLAs), robots can unify visual perception, language understanding, and physical action in a single computational framework. In February 2025, Figure AI released Helix, a VLA model for generalist humanoid control. Helix sets new standards in the VLA and robotics space by enabling zero-shot generalization, system1/system 2 architecture. With zero-shot generalisation, no extensive retraining is needed for each task the robot does. Helix can generalize instantly to new situations, objects, and instructions. Under system 1/ system 2 architecture, high-level and lightweight reasonings are separated, enabling commercially viable humanoids that combine human-like reasoning with real-time precision.
Affordable Robots are Visible
Technologies that change the world all have one thing in common – they are accessible. Smartphones, personal computers, 3D printing are all made accessible through being affordable by the middle class. When robots such as the Unitree G1 are cheaper than a Honda Accord or the minimum wage of 34K USD in the States, it is not hard to imagine a world where physical and mundane tasks are mostly done by robots.
Source: https://www.unitree.com/g1/
Expanding from Warehouse to Consumer
Robotics is expanding from warehouse solutions to consumers. The world is made for humans. Humans can do everything specialised robots can; specialised robots can’t do everything humans can. Instead of making specialised robots for factories, robotics companies are making more generalised humanoid robots. Thus, the frontier of robotics can not only be observed in warehouses, but in our everyday lives.
Affordability is one of the main bottlenecks for expansion. The metric I look at the most is cost per hour. We define the cost as the opportunity cost of time used for training and charging, the cost of doing the task and the cost of the robot divided by its operational hours. The average wage of the sector is the benchmark to go under.
Source: https://www.bls.gov/news.release/empsit.t19.htm
In order to fully penetrate the warehouse sector, cost per hour must be lower than 31.39 USD. The largest consumer sector above is Private Education & Health Services. Robots must cost less than 35.18 USD per hour to penetrate consumer verticals.
Robots are becoming cheaper, more efficient and more generalised.
What are the Next Steps for Robotics?
Battery Optimization
Battery has always been a bottleneck for user-friendly robots. Early electric vehicles such as BMW i3 struggled to gain popularity due to weak battery technology that limited range, increased cost, and reduced practicality. Robots are facing the same challenges. Boston Dynamics’ Spot robot can only operate for 90 minutes before needing a recharge; Unitree G1 has a battery life of around 2 hours. No one wants to manually recharge their robots’ batteries every 2 hours. Therefore, autonomous recharge and docking infrastructure is an area worth looking at. Currently there are two ways robots are recharged – swapping the battery or charging the battery.
Swapping battery trays allow depleted batteries to be quickly swapped out for charged ones, minimizing downtime and enabling robots to operate continuously in the field or on the factory floor. This process can be done both manually and automatically.
Inductive charging charges robots wirelessly, it takes more time to fully recharge the robot but the process can easily be fully automated.
Latency Optimization
Low latency operations can be divided into two categories: perception and teleoperations. Perception is how the robot perceives its surroundings. Teleoperation is when a human is remotely controlling the robot.
According to Cintrini Research, robotic perception starts with cheap sensors but the moat forms in fusion software, low-power compute, and millisecond-tight control loops. Once the robot knows here and there, a lightweight neural network tags obstacles, pallets, or humans. Scene labels feed the planner, which then spits out motor commands to feet, wheels, or arms. Perception latency under 50ms is equivalent to a human’s reflexes - anything slower and your robot is going to be really clumsy. Thus, 90% of the decisions live onboard in the form of a single vision-language-action network.
For a fully autonomous robot, the latency of the high performant VLA must be lower than 50ms; for a teleoperator robot, the latency between the operator and the robot must be lower than 50ms. We can fully appreciate the importance of VLAs here, if visual and text inputs are processed in two models separately and fed into a larger LLM for processing, the latency would be way higher than 50ms.
Data Collection Optimization
There are three ways to collect data: real world video data, synthetic data, teleoperation data. The main constraints for real world data and synthetic data is bridging the difference between how robots behave in physical settings versus how they are modeled in video or simulations. Real world video data lack physical details such as force feedback, joint movement inaccuracies, or material deformation; Simulation data lack unpredictable variables such as sensor failure and friction.
The most promising data collection method is teleoperation, where human operators control robots remotely to perform tasks. Capital for human labor is the major constraint for teleoperation data.
Custom hardware is also being developed to provide high quality data as well. Mecka is providing high-quality, high-volume human movement data with rapid iteration cycles for AI robotics training. It captures several forms of human data with both mainstream methods and custom hardware that is then processed and transformed to be usable in training robotics neural networks. Together, these pipelines shorten the path from data to deployable robotics.
Things to Look at in the Space
Crypto Meets Robotics
Crypto can be used to incentivize trustless parties in improving network efficiencies for robotics. Looking at the key areas above, we believe that crypto can enhance efficiencies in docking infrastructure, latency optimization and data collection.
Decentralized Physical Infrastructure Networks (DePIN) has the potential to docking infrastructure. When humanoid robots, like cars, travel around the world, it is important for charging stations to be as accessible as gas stations. Centralized networks require heavy upfront investment, while DePIN distributes the cost among node operators, allowing charging infrastructure to expand quickly and reach more locations.
DePINs are also positioned to leverage decentralized infrastructure for optimizing latency in teleoperations. DePINs can aggregate computing resources from geographically dispersed edge nodes, teleoperations from. This means commands from a teleoperator to the robot can be processed locally or by the nearest available node, minimizing the distance data must travel and potentially reducing communication latency. However, current DePIN projects primarily focus on decentralized storage, content delivery, and bandwidth sharing. Some projects have demonstrated the benefits of edge computing for applications such as streaming or IoT, but not in robotics or teleoperations.
Teleoperation is the most promising data collection method. However, it is very expensive for a centralised entity to hire specialised personnel to produce teleoperation data. DePIN solves this by incentivising third parties in providing teleoperation data. Reborn uses crypto tokens to reward and coordinate a global network of teleoperators. It turns their contributions into tokenized digital assets, enabling a decentralized, permissionless system where anyone can earn, govern, and help train AGI robots.
Safety is Always a Concern
It is common knowledge that the end goal for robotics is fully autonomous robotics. However, as seen in the block buster Termaintor, the last thing humanity wants is when their autonomy turns them into aggressive robots. AI safety has been a concern for LLMs. But when LLMs have limbs, robotics safety is crucial for robotics to be adopted in society.
Economic security is one of the key pillars for a flourishing robot economy. OpenMind, a company in the space, is building FABRIC—a decentralized coordination layer for machines to establish identity, verify physical presence, and access capital or jobs through cryptographic proofs. Instead of simply managing task marketplaces, FABRIC enables robots to prove who they are, where they are, and what they’ve done—without relying on centralized intermediaries.
Behavioural guardrails and identity attestations are enforced on-chain, allowing anyone to audit compliance. Robots that meet safety, quality, and location criteria are rewarded; those that fail are slashed or disqualified, ensuring both accountability and trust in autonomous machine networks.
Third parties restaking networks such as Symbiotic can also provide equal security guarantees. There is still work to be done on slashing parameters but the technology is already in production. We believe industry wide safety guidelines will be developed soon and slashing parameters would be modelled after the guidelines.
An example of how this can be implemented:
A robotics company joins Symbiotic as a network
Verifiable Slashing parameters such as “Making physical contact with humans at the force exceeding 2,500 newtons” are established
Stakers stake to provide guarantees that the robot wouldn’t violate the slashing parameters
If the robots violated the slashing parameters, the stake would be used as compensation to the victim
Under this model, robotics companies are incentivised to put robotics safety as their number 1 priority, while the insurance pooled by the stakers would enhance the consumer adoption of robotics.
Here’s what the Symbiotic team thinks about robotics:
Symbiotic’s universal staking framework was designed to extend the staking concept to any vertical or protocol that could benefit from economic security, whether through shared or individual models. Applications range from insurance to robotics and need to be conceptualized on a case-by-case basis. Robotic networks, for instance, could be built exclusively with Symbiotic’s universal staking framework, enabling stakeholders to provide economic backing for network integrity.
Filling Holes in the Robotics Stack
OpenAI popularised AI, but the groundwork for the ChatGPT moment was done years in advance. Cloud services made it possible for models to not be constricted to local compute, Huggingface made it possible for models to be open-sourced, Kaggle provided a space for AI engineers to experiment in the space. All these baby steps made it possible for AI to be popularised.
Unlike AI, It is hard to get started in robotics with limited capital. For robotics to be popularised, building robots should be as easy as building AI applications.
We believe there is room for improvement in three layers – financing, evaluation and education.
Financing is an issue for robotics. You only need a computer and cloud computing credits to build a computer. To build a fully functioning robot, you need to buy hardware such as motors, sensors and batteries, the cost of which can easily be north of 100K USD. The hardware nature of robotics makes building robotics less flexible and more expensive than building in AI.
The infrastructure for real world robotics evaluation is nascent. In AI, loss functions are well defined and testing operations can be done virtually. However, a good virtual policy doesn’t translate directly to a good policy in the real world. Infrastructure for evaluation of autonomous policies in diverse real world environments is needed for robots to iteratively improve.
Once these rails solidify, talent will flood in and autonomous humanoids will follow the same adoption curve that catapulted Web 2. OpenMind, a crypto robotics company, is already pushing in that direction. OM1, the company’s open-source “Android for robots”, turns raw hardware into upgradeable, economically aware agents. Vision, language, and motion planners click in like phone apps, and every reasoning step appears in plain English so operators can audit or redirect behavior without touching firmware. This ability of reasoning in natural language empowers the next generation of talent to seamlessly move into robotics, making for a great first step to the kind of open platform that can spark the robotics boom the way open source accelerated AI.
Source: A humanoid robot powered by OpenMind’s OM1, the world’s first open-source, AI-native operating system for robots, pressed the button to celebrate an ETF listing on the NASDAQ.
Talent density defines an industry’s trajectory. Structured and accessible education is crucial for talent to be funneled into robotics. OpenMind’s Nasdaq appearance signals the beginning of a new era where intelligent machines play a direct role in both financial innovation and real-world education. OpenMind and Robostore announced the rollout of the first widely adopted educational curriculum for the Unitree G1 humanoid robot across K–12 public schools in the United States. The curriculum is designed to be platform-agnostic and adaptable to various robotic form factors, offering students hands-on experience with robotics. This is an encouraging sign and strengthens our belief that the amount of education material for robotics will be on par with that of AI in a few years.
Looking Ahead
VLA innovations and economies of scale have catalyzed the creation of affordable, efficient and generalised humanoid robots. Robotics safety, financing and evaluation are worth exploring as warehouses expand into consumer robotics. We also have strong conviction that crypto will improve robotics by providing economic guarantees for robot safety, and optimising its docking infrastructure, latency and data collection pipelines.
- Paul Veradittakit
Business
Centrifuge joins Converge as a launch partner
Converge names Centrifuge as its first launch partner adding institutional grade RWA products and developer tooling to its new tokenized marketplace.
U.S. firms pour $844 mln into crypto treasuries
American companies moved $844 million into Bitcoin and Hyperliquid (HYPE), driving a jump in open interest and bullish sentiment for both tokens.
Regulation
Fannie Mae, Freddie Mac ordered to consider crypto as an asset when buying mortgages
FHFA Director William Pulte has ordered Fannie Mae (FNMA) and Freddie Mac (FMCC) to consider cryptocurrency as an asset for single-family mortgage loan risk assessments.
Coinbase Secures MiCA Licence: A Milestone in Europe’s Crypto Evolution
Securing registration in Luxembourg lets Coinbase passport its full suite of crypto services across all 27 EU states under the new MiCA regime.
GENIUS Act Stablecoin Bill Passes Senate
The U.S. Senate approved the GENIUS Act with bipartisan support. The bill mandates full reserves and regular, audited disclosures for large stablecoin issuers, along with clear redemption rights and operational standards.
New Products and Hot Deals
Zama unveils the Zama Confidential Blockchain Protocol and announces their $57m Series B
Zama is unveiling their most ambitious product to date: the Zama Confidential Blockchain Protocol, and announcing their $57m Series B at a valuation of over $1b. The round was led by Pantera Capital and Blockchange.
Stripe acquires Privy: Bringing crypto to everyone
After the recent $1.1b acquisition of Bridge, Stripe is now buying Privy and integrating its 75 M self-custodial wallets into Stripe’s payments stack while keeping the Privy product independent.
Nexus rolls out TestNet III
The verifiable-compute blockchain launches its TestNet across the United States, Vietnam, Nigeria, Russia, India, Indonesia, the United Kingdom, Hong Kong SAR and Mainland China.
This Week at Pantera
Pantera partners with Blockworks Research on the Token Transparency Framework
Cosmo Jiang and Eric Wallach explain consistent disclosure standards for token-based protocols.
Dan Morehead on Bloomberg - Trading Cryptocurrencies Like Hard Assets
Pantera’s Dan Morehead tells Bloomberg’s “Open Interest” that institutional traders now treat Bitcoin like gold, using derivatives and robust custody to hedge geopolitical shocks. Favorable regulations unlock fresh fundraising.
World: A Mission Critical Identity Solution
Pantera believes World’s proof-of-human solution will become critical infrastructure in a future where humans and AI coexist.
ABOUT ME
Hi, I’m Paul Veradittakit, a Managing Partner at Pantera Capital, one of the oldest and largest institutional investors focused on investing in blockchain companies and cryptocurrencies. I’ve been in the industry since 2014, and the firm invests in equity, early-stage token projects, and liquid cryptocurrencies on exchanges. I focus on early-stage investments and share my thoughts on what’s going on in the industry in this weekly newsletter.
If you have any projects that need funding, feel free to DM me on twitter.
Very interesting article I grew up with robotics and now I’m heavily in crypto I love this 🏆