Highlights
- Reduced average inference latency from 9.03 seconds to 3.98 seconds.
- Achieved more than 2× higher throughput.
- Enabled benchmarking and scalable testing of open-source LLMs for industrial robotics.
Challenges
ARENA2036 stands for “Active Research Environment for the Next generation of Automobiles” and is one of nine research campuses of the “Research Campus – Public-Private Partnership for Innovation” funding initiative in Germany. ARENA2036 is supported by the Federal Ministry of Research, Technology and Space and is run as a registered association with members from science and industry. These members are active in various disciplines – from the automotive industry to aerospace technology, textile and materials research to ergonomics. Programming industrial robots traditionally relies on manually designed rule-based systems that are difficult to scale. Large Language Models offer a new paradigm by translating natural-language instructions into executable robot actions, but evaluating hundreds of planning tasks requires substantial computational resources.
Solution
Through HammerHAI, ARENA2036 leveraged the HLRS Hunter supercomputer to benchmark an LLM-based robot task planning framework using DeepSeek-Coder. Hunter enabled large-scale evaluation across 90 manipulation tasks comprising 450 robot actions, providing the computational performance required for systematic benchmarking and optimization. HammerHAI provided the computational infrastructure that transformed experimental AI concepts into scalable industrial solutions.
Benefits and Impact
Using HLRS Hunter, average inference latency was reduced from 9.03 s to 3.98 s, with more than 2× higher throughput across a benchmark of 90 tasks and 450 robot actions. HammerHAI enables scalable evaluation of LLM-based robotic intelligence without requiring companies to build dedicated AI computing infrastructure. Through AI-MATTERS, these approaches can be tested in realistic industrial robotics environments before deployment.
Organizations Involved
Contact
Muhammad Saeed, ARENA2036, muhammad.saeed@arena2036.de
