Probabilistic Trajectory Reconstruction from Keyframes

Status: Ongoing

Reconstructing high-frequency robot trajectories from sparse keyframes with calibrated uncertainty. Compared per-joint Gaussian Process baselines against a masked variational autoencoder on a real SO-101 manipulator dataset — the VAE wins at sparse keyframe rates (≤ 1 Hz) by exploiting dataset-wide motion structure; the GP stays competitive at dense rates.