systems SHARD End-to-end VLM inference on the Rockchip RK3588 NPU OpenNPU An open compiler and runtime for the RK3588 NPU FP8-as-Storage GEMM for Ampere FP8 matmul on an RTX 3090, no H100 required robotics LIBERO on MuJoCo Warp 130 robot manipulation tasks on one $700 GPU Neuro-Symbolic VLM for Adaptive Impedance Control Connecting a VLM's scene understanding to a low-level impedance controller Compliant Explicit Reference Governor Provably safe contact for robotic manipulators