User Goals
- Anticipate future scene graphs and process states from egocentric industrial video.
- Fine-tune compact multimodal models that output validated, machine-readable results.
- Enable real-time planning and decision-making for robotics and manufacturing.
Current State and Key Challenges
Ramblr GmbH is a German AI startup that builds models that understand the physical world to support robotics and industrial production. Our technology turns egocentric video (i.e., video shot from an individual subject’s perspective) into structured representations of objects, attributes, spatial relations, activities, and process states. A new technology currently under development at Ramblr—called JAWS—goes further, anticipating plausible short-term futures as scene graphs in validated, machine-readable form.These graphs can then be used for downstream planning.
We train with a production-tested stack based on PyTorch FSDP, Ray, and containerized workflows for distributed GPUs. However, we currently face the challenge of scaling. Fine-tuning multimodal models on industrial video data requires substantial computing power, fast connections between computing resources, reliable storage, and repeatable execution across multiple nodes. In addition, industrial data are not consistent across sites, processes, camera positions, and conditions. Models must learn robust representations while producing outputs in a format upon which machines can act. For a startup, limited GPU access makes systematic experimentation and validation difficult. HammerHAI provides the infrastructure to move from smaller experiments to production-relevant training.
Support from HammerHAI
Through HammerHAI, we have been using scalable NVIDIA GPU infrastructure, high-performance networking, and storage for distributed multimodal training at KISSKI (AI service center for critical and sensitive infrastructures) in Göttingen. Technical onboarding and support for efficient multi-node execution have helped us to validate our FSDP, Ray, and container-based pipeline at production-relevant scale.
Intended Value and Impact
With JAWS we aim to deliver a compact multimodal world model that predicts scene graphs and process states from industrial video in near real time, targeting inference within 200 milliseconds per frame. Machine-readable outputs will help robots and assistance systems recognize production situations, anticipate what happens next, and react more reliably to change. Our goals are to shorten development cycles, improve robustness across industrial environments, and build a scalable foundation for AI-enabled manufacturing.
Current Status within HammerHAI and Next Steps
JAWS has been approved for the HammerHAI training platform provided through KISSKI (AI service center for critical and sensitive infrastructures) and GWDG (Gesellschaft für wissenschaftliche Datenverarbeitung mbH Göttingen), with a project period from September 2026 to April 2027. As next steps we will onboard software to the GPU cluster, validate our distributed environment, transfer representative training data, and run initial scaling benchmarks to configure fine-tuning and evaluation.
Contact
Philipp Schubert, Ramblr GmbH, philipp@ramblr.ai
