CDPs add a simple contraction regularizer to diffusion policies, pulling nearby sampling trajectories together to suppress solver and score-matching errors. This yields more robust action generation in offline RL and imitation learning, especially in low-data regimes.
From a low-level Cortex API to a full ROS2 bringup with real/sim parity, teleop, calibration, and fault handling—this is the Link6 infrastructure I built to make learning on a real cobot actually possible.
VOCALoco predicts the viability and cost of transport for several pretrained locomotion skills from local heightmaps, then executes the safest, most efficient one. It improves robustness on stairs and transfers zero-shot to real hardware.
DexAloha is my build of the open-source HOPE JR hand: I assembled it, designed a custom motorized wrist that plugs into a Kinova Jaco, and (with Martino Russi) adopted an improved finger design that is easier to assemble and route cables through.
The long story of tracking a one‑minute‑then‑die power fault, opening smart batteries, rebuilding the base from first principles, adding modern sensing/compute, and restoring dual‑arm manipulation and giving a demo to the CEO of Kinova.
How I built a common Isaac Sim setup for MOVO and my Jaco table, cleaned up URDFs, aligned ROS2 interfaces, and created a shared playground where the same policies can run in sim and on the real hardware.
How I designed and built a custom robot arm environment with a sturdy frame, cage, power management, kill switch, sensors, and dual Kinova Jaco arms for real-world experiments.