I ran 130 robot manipulation tasks in parallel on a $700 GPU
LIBERO, a 130-task robot manipulation benchmark, ported to MuJoCo Warp and run on one RX 9070 XT, at 17,137 physics env-steps/s, with a BC transformer at 42.5% and the four renderer fixes it took to get there.
OpenNPU v1.0: an open compiler and runtime for the RK3588 NPU
OpenNPU v1.0 is an MIT-licensed compiler and runtime that runs ONNX, PyTorch, and JAX models on the Rockchip RK3588 NPU through raw DRM ioctls, with no closed-source dependency. This report covers the hardware contract, the register-command wire format, the compiler pipeline, and the measured results.
Opening the black box: a year building an open compiler for the RK3588 NPU
A year on the Rockchip RK3588 NPU: from hacking a closed toolchain to run one vision encoder, to an open, extensible compiler and runtime that works from PyTorch, ONNX, and JAX.
Repairing a Frozen Visuomotor Policy With Counterfactual Regret Labels & Flow Matching
Teaching a frozen BC policy when to ask for help, and how to push its actions back toward expert behavior using counterfactual regret labels and flow matching in DCT space.
Paper Review: No Plan but Everything Under Control
How one ICRA 2025 paper replaces planning, state machines, and learned policies with plain gradient descent — and it actually works.