Robot skill adaptation typically relies on reinforcement learning (RL) refinements, but existing skill models often struggle beyond minor environmental changes. Gaussian process (GP)-based skill modelling offers expressive representations with useful analytical properties, yet adapting GP-based skills has remained underexplored. This paper proposes a robust adaptation framework that uses GP models with sparse via-points to represent trajectories compactly while preserving the skill's kinematic profile through first and second derivatives.

The authors present three adaptation methods to handle variability between initial and observed configurations. The first is an optimisation agent that adjusts via-points while maintaining demonstration velocity. The second is a behaviour cloning agent trained to replicate the optimisation agent's output trajectories. The third is an RL agent that modifies via-points online while preserving the kinematic profile.

Evaluated across three tasks—drawer opening, cube pushing, and bar manipulation—the methods outperformed every benchmark in success rates in both simulation and hardware. The GP-based representation also enabled all three methods to achieve high cosine similarity and low velocity magnitude errors, indicating strong preservation of the kinematic profile. The formulation provides a compact representation capable of adapting to large deviations from a single demonstrated skill.