.png&w=256&q=75)
🤓 Latest Submissions
Percept2act
percept2act is an entry for the Intel Physical AI Challenge. An SO-101 robotic arm inspects a LEGO brick at a fixed pose, decides whether it is defective, and places it in one of two plates. Defect detection uses Anomalib PatchCore, fitted on 90 normal crops and exported to OpenVINO IR. It runs on the NPU of an Intel Core Ultra Series 3 at 51 ms per inference. Ten crops of each class were withheld from every fitting and tuning decision, and the model classifies all of them correctly. The verdict then selects one of two English sentences. That sentence conditions a SmolVLA policy fine-tuned on 40 teleoperated episodes, so one checkpoint routes a brick to either plate based on text alone, with no branching logic in between. Work is divided across all three engines. The detector runs on the NPU, the policy on the Arc integrated GPU, and the control loop on the CPU. Every stage logs its own device and latency.
16 Sep 2026
