Accelerating Research in Autonomous Driving
Toyota Research Institute
SUMMARY
Toyota Research Institute (TRI) is doing cutting-edge R&D work in the autonomous driving space.
TRI needed advanced engineering and operations support to scale its machine learning (ML) research efforts.
Hop has worked as an embedded part of TRI’s team, building the infrastructure to accelerate their work.
Our collaboration has resulted in deployment in TRI research vehicles, as well as time and cost savings for TRI.
THE COMPANY
Toyota Research Institute (TRI) stands at the forefront of innovative research and development in the field of autonomous driving and human-vehicle interaction. With a focus on real-world application and a reputation for groundbreaking research, TRI is a trailblazer in shaping the future of mobility. Hop Labs is proud to have established a long-standing and trusted partnership with TRI’s Human Interactive Driving (HID) division. Their work aims to create a seamless and intuitive relationship between humans and self-driving vehicles, ensuring that these autonomous systems work to amplify human drivers for a safer and more enjoyable driving experience.
THE CHALLENGE
At the time we connected with TRI, their HID research team faced challenges as their experiments grew in complexity and scale. Engineering needs of the researchers outpaced their existing capabilities and experimental datasets were approaching hyperscale levels, far surpassing the capabilities of off-the-shelf products. Training was hindered by diverse data formats and technical debt. Compute costs and resource allocation were becoming more of a concern, as well as the need for integration of separate lines of research within TRI. Recognizing the need for advanced engineering and operations support, they turned to Hop Labs.
“The Hop team was very proficient and helped us achieve our internal milestones, with strong data/ML capabilities and agility to handle multiple research threads, supercharging our R&D team.”
THE APPROACH
Since 2020, Hop has supported TRI’s research efforts and accelerated the path toward valuable insights and significant outcomes. This journey has been characterized by a deep understanding and close alignment of goals, facilitated by our engineers working as embedded team members within TRI’s research team.
Our teams have collaborated closely to design and construct pipelines capable of handling datasets on the order of petabytes. In supporting TRI’s ML researchers, Hop has provided not only engineering expertise but also assistance with cloud infrastructure and data access to accelerate operations. We’ve maintained a focus on optimization to increase efficiency and cost savings, balanced with an understanding of where flexibility and investment were needed to support TRI’s research.
THE RESULTS
Hop has empowered TRI with a range of solutions. We’ve built a large-scale tool that effectively handles the ingestion, indexing, retrieval, and visualization of vast amounts of data. This platform now allows researchers to seamlessly build datasets across diverse formats and easily extract actionable insights. By scaling cloud compute infrastructure, the tool has also significantly reduced data processing time from several days to just a few hours.
The HID team’s training process has been accelerated through profiling and caching, resulting in a reduction in training time from several weeks to two days. The training framework is now both robust – with comprehensive testing, improved GPU utilization, and a suite of visualization tools – and flexible, supporting new use-cases and model architectures.
Implementation of efficient cloud computing solutions has been key for managing costs and expediting research for TRI. Reducing idle time and employing lower-cost resources has saved tens of thousands of dollars per month. Quick provisioning has reduced experiment startup time from an hour to a minute, and deployment of an optimized compute cluster has enabled a four-fold increase in the number of experiments, all without increasing cloud costs.
Our TRI researcher counterparts now enjoy an expanded level of freedom in designing experiments, as they are no longer constrained by limitations of dataset size or expensive cloud compute. Additionally, we’ve integrated autonomous driving models with human-facing simulators, advancing capabilities and opportunities across teams within the organization. Hop is proud to be part of real-world impact created by the cutting-edge research in our partnership with TRI.
Does your team need additional ML engineering capacity? Are you looking for ML operations support? Contact us to learn how Hop can help.