The research runs in two directions: making models small enough to run on hardware that cannot afford them, and making robotics software stop caring which hardware it is on.
Peer reviewed and available through the venue.
arXiv preprint, 2026
Presents the Meta-ROS middleware architecture, cross-platform communication layer, and benchmarks against ROS 1 and ROS 2, reporting up to 30 percent higher throughput than ROS 2.
CDSR 2025
Surveys the robotics middleware landscape and the compatibility problems that motivated Meta-ROS, comparing transport models, message schemas, and platform support across existing systems.
IEEE MoSICom 2024, Dubai
Fuses a classical HMM front end with a deep acoustic model to make speech recognition viable on embedded robot hardware, reaching up to 26.59x faster inference on a Jetson Xavier NX.
Project paper, 2024
Implementation report for a CUDA-kernel-free RWKV language model that combines linear attention, local attention, and mixture-of-experts layers in Python and PyTorch.
Citations, BibTeX, and the reviewer correspondence for any of these are available on request. The MicroRWKV write-up on the work page includes the loss table exactly as the training runs produced it, including where the model comes in behind its baseline.