Hey HN! I'm Lakshay. Raveen and I have built Cerulion, a ROS2-compatible robotics runtime written in Rust. We've worked hard on lowering IPC transport latencies without making the single process tradeoff or forcing POD message types. These fast multi-process graphs can then be resimmed deterministically, and existing ROS2 nodes/launch files can run unmodified alongside native Cerulion Rust/Python code. Installation and quickstart (with a one-shot prompt for your favorite coding agent) here at the repo README: https://github.com/cerulion-inc/cerulionI worked on off-road autonomy during my PhD at MIT, and we kept running into inter-process latency in ROS (2 lidars and 6 cameras). Loaned messages need POD types, which nukes standard, vendor-agnostic Image/PointCloud2. We ended up running some hot paths in a brittle single process, which made debugging a nightmare. I'm sure you've seen the heisenbugs ROS-based stacks can produce, with no way to easily reproduce the consumption order or interleaving of callbacks. Valuable data of runtime failures on the robot was hard to use effectively without deterministic reproducibility.This is what we fix. Cerulion bags record per-input reads and scheduler traces alongside message content in standard MCAPs. cerulion bag play --resim all --verify re-executes node code on the recorded schedule and compares each byte for exact matches or within defined tolerances. Exit 0 is a match, 1 shows the first topic/frame that diverges, 6 shows schedule divergence and --report gives an optional CI verdict. Eligible ROS2 processes can replay deterministically as well.Here's the data loop:1. Run: Dependency-ordered level-based DAG scheduler with periodic/data/sync triggers, I/O deadlines, backpressure, and automatic profiling for process partitioning. Each node by default free-runs in its own process while preserving deterministic replay. Live runs pause/resume, and UMWAIT/WFE parks idle consumers on CPUs that support it. Wall-clock reads and runtime thread spawns in user code rejected at compile time by macros.ROS 2 is supported with rmw_cerulion for Foxy, Humble, Jazzy and Lyrical, that lets you run ROS2 nodes/launch files unmodified in the same graph as native Cerulion nodes. LD_PRELOAD gives ROS2 code shared memory loans through std::vector/std::string. cerulion ros2 attach folds in already-running ROS2 robots, and cerulion ros2 run/launch lets you run existing ROS2 workspaces on Cerulion without any additional setup (with an optional automatic migrate suggested that rewrites C++ publish sites).2. Move: iceoryx2 shared memory with no serialization, no payload copy and no payload-sized allocation, including for variable-length messages like sensor_msgs/Image and PointCloud2. Round trip latency for a 1MiB sensor_msgs/Image between two processes is 4.08µs p50, with a tight tail of 7.38µs p99 (under 2x p50),