ROSA

ROSA (Robotics Oriented Serving Architecture) is a serving system for Robotics Foundation Models (RFMs) designed for robot factories. It is built on three principles: shared GPU-pool serving where a fleet of robots accesses server-class GPUs over the network, a robotics-aware programming abstraction supporting multi-model pipelines with per-task SLOs and failure handling, and factory-objective-driven scheduling. Implemented on Ray Serve with vLLM, PyTorch, and JAX backends.