Associate Prof. Takamitsu Matsubara (Robot Learning Laboratory)
Bayesian Disturbance Injection: Robust Imitation Learning of Flexible Policies
In this talk, we discuss imitation learning problems for learning robot skills from human demonstration data. Scenarios requiring humans to choose from multiple seemingly optimal actions are commonplace, however standard imitation learning algorithms often fail to capture this behavior. Instead, an over-reliance on replicating expert actions induces inflexible and unstable policies, leading to poor generalizability in an application. To address these problems, we introduce our imitation learning framework that incorporates Bayesian variational inference for learning flexible non-parametric multi-action policies, while simultaneously robustifying the policies against sources of error, by introducing and optimizing disturbances into expert actions to create a richer demonstration dataset. This combinatorial approach forces the policy to adapt to challenging situations, enabling stable multi-action policies to be learned efficiently. The effectiveness of our proposed method is evaluated through simulations and real-robot experiments.