How io.net Cut AI Research Costs 92.8% for Frodobots-UC Berkeley Breakthrough

Frodobots Challenge: Proving their crowdsourced data could power breakthrough navigation AI through reliable AI infrastructure

Key Results:

• 92.8% cost savings compared to Amazon Web Services’ H100 pricing
• 12,696 GPU hours across 8 GPUs with zero failures
• 66-day project duration with seamless deployment
• 5TB storage provisioned in 30 minutes vs. days with Big Tech cloud providers
• AI infrastructure reliability that led to 85.7% navigation success

"We chose io.cloud because it delivered reliability when other platforms couldn't. When we needed 5TB of additional storage for our massive datasets, the io.net team provisioned it in 30 minutes. That's what sealed the deal for us." — Catherine Glossop, UC Berkeley RAIL Lab

Overview

Frodobots, a Singapore-based AI robotics startup, transformed their approach to data validation by partnering with UC Berkeley Robotic AI & Learning (RAIL) Lab and io.net to prove their crowdsourced navigation dataset could power breakthrough AI research. Initially struggling to demonstrate the value of their 2,000+ hour navigation dataset (25 times larger than any competitor), Frodobots needed reliable machine learning infrastructure to support rigorous academic research that would validate their data assets.

Through io.cloud's on-demand, high-performance GPU cloud computing platform, the collaboration successfully processed 12,696 GPU hours, including a continuous 10-day training run across 8 GPUs without failures. UC Berkeley's research achieved an 85.7% navigation success rate compared to just 33.3% for baseline methods. The partnership resulted in a peer-reviewed research paper, global validation across 6 countries, and positioned Frodobots as a leader in embodied AI research.

The Challenge

About Frodobots

Frodobots raised $8M in funding from investors including Protocol VC, Solana Ventures, and Solana co-founders. The 12-person team has built an innovative "robotic gaming" platform where players earn rewards by remotely controlling real robots to complete navigation missions worldwide.

Frodobots operates robots across multiple cities, collecting extensive navigation data from real-world deployments. The company had accumulated over 2,000 hours of valuable navigation data from 10+ cities worldwide - a dataset 25 times larger than other publicly available navigation datasets. However, they faced a critical challenge: how to demonstrate this dataset's value for advancing AI research.

The Strategic Problem

Traditional data licensing wasn't enough. Frodobots needed to prove their dataset could train breakthrough navigation models that work in any environment, but academic researchers consistently struggled with compute limitations. Without adequate research infrastructure, even the most valuable datasets couldn't reach their full potential.

The company recognized that to establish thought leadership and validate their data assets, they needed to move beyond simple licensing. They had to invest in enabling world-class research that would definitively prove their dataset's value for training generalist navigation models.

The Solution

Why Frodobots Chose io.net's AI Infrastructure

When Frodobots and UC Berkeley's RAIL Lab assessed what they would require to execute on their ambitious research project, they realized they needed AI infrastructure that could handle the demands of cutting-edge research. The project required processing 2TB of navigation data through week-long training runs without interruption.

Frodobots provided their unique navigation dataset, UC Berkeley RAIL Lab conducted the research, and io.net sponsored the GPU cloud computing infrastructure to make breakthrough research possible.

But the RAIL Labs team needed more than just raw compute power. When the research required additional storage to handle massive datasets, io.net's technical team quickly provisioned over 5 terabytes of additional local storage in 30 minutes - compared to days with traditional cloud providers.

The Implementation

From Feb 24 to May 1st, 2025, UC Berkeley RAIL Lab used dedicated H100 nodes through Frodobots' partnership with io.net for 66 days straight. io.cloud's machine learning infrastructure supported 12,696 total GPU hours, including one continuous 10-day training run across 8 GPUs without interruption while processing 6,000 hours of trajectory data.

The integration process proved remarkably smooth. As Glossop noted, "Transferring our PyTorch-based codebase to the io.net server was straightforward. The dedicated AI infrastructure approach proved far superior to alternatives. The setup was like having a dedicated machine in our lab specifically for our project, which was a welcome relief from the hassle of borrowing resources from other cloud platforms."

Beyond compute power, io.net's responsive technical support became crucial when research needs evolved. When the team required additional storage capacity to handle their massive datasets, io.net's technical team provisioned over 5TB of additional local storage in just 30 minutes - a process that typically takes days with Big Tech cloud providers.

The infrastructure delivered consistent performance throughout the entire 66-day project timeline, enabling researchers to focus on their work rather than managing technical bottlenecks or dealing with the preemption issues common in shared cloud environments.

The Results

AI Infrastructure That Enabled Research Breakthrough

The partnership delivered exactly what Frodobots needed: the infrastructure reliability to enable breakthrough AI research. UC Berkeley RAIL Lab successfully published a peer-reviewed research paper demonstrating how Frodobots' navigation data could train generalist navigation models - research that was only possible because of io.cloud's on-demand, high-performance GPU cloud computing capabilities.

As Frodobots CEO Michael Cho explains, "Having UC Berkeley RAIL Lab, one of the world's top robotics research institutions, spend significant time publishing peer-reviewed research that produced state-of-the-art results with our dataset definitely helped publicly validate and add legitimacy to Frodobots' mission." This breakthrough was enabled by io.cloud's ability to provide uninterrupted training sessions and rapid storage scaling capabilities that Big Tech cloud providers and university resources couldn't match. io.cloud enabled validation across 6 countries spanning 3 continents, providing the consistency needed for global-scale testing.

Technical Validation

The AI infrastructure delivered exactly what the research demanded: 12,696 GPU hours, including a 10-day continuous run without a single failure. Glossop emphasizes how io.net's approach differed: "Compared to other cloud compute providers like AWS or Google's TPU Research Cloud, working with io.net was extremely smooth. We had access to a consistent, dedicated machine rather than dealing with virtual machine instance setups that could get preempted."

The ease of onboarding with io.net was just as crucial for the RAIL Labs team. "Transferring our PyTorch-based codebase to the io.net server was straightforward. It was like having a machine in our lab specifically for our project," Glossop notes. This reliability was key for the demanding computational requirements, as she explains: "With our lab's machines constantly overloaded and cloud resources being unreliable, io.net's dedicated approach let us process 2TB of navigation data and complete week-long training runs that would have been impossible elsewhere."

Key Performance Insights

• Speed: 5TB storage in 30 minutes vs. days elsewhere‍
• Reliability: 12,696 GPU hours without preemption vs. constant interruptions‍
• Cost: 92.8% cost savings vs. AWS pricing

This machine learning infrastructure advantage has translated into concrete business results for Frodobots. As CEO Michael Cho explains, "We actually have about a dozen ongoing research collaborations with other university labs at the moment, but having the first one with UC Berkeley set the stage for the rest, and gave teeth to our positioning as a world-class collaboration partner."

Beyond immediate research outcomes, the partnership demonstrates how modern AI infrastructure can accelerate academic breakthroughs. As Cho notes, "Having io.net involved in this collaboration has been an important step towards showing researchers that crypto, when done well, can really move the needle for them and help them to do their best work."

About This Partnership

This case study showcases how AI startups can leverage strategic compute partnerships to validate their data assets and establish academic credibility. Had Frodobots used AWS pricing, the compute would have cost $144,735 more (92.8% premium), demonstrating how efficient GPU cloud computing partnerships can make rigorous academic research economically viable while delivering breakthrough results.

Ready to accelerate your AI research partnerships?

Try io.cloud today or contact our team to discuss your infrastructure needs.