Ludovic Righetti

Robots can now compute collision-free, visually conditioned trajectories directly on standard 3D Gaussian Splats at high throughput without ever converting the scene into lossy meshes or voxel grids.

A single, unified cost function, learned via inverse reinforcement learning, accurately predicts human reaching movements across subjects and postures, suggesting a general optimality principle.

Hippo offers a robust and efficient alternative to SQP and DDP for robotic trajectory optimization, particularly for problems with hard constraints.