Scalable and Efficient Continual Learning from Demonstration via a Hypernetwork-generated Stable Dynamics Model

Learning from demonstration (LfD) provides an efficient way to train robots.The learned motions should be convergent and stable, but to be truly effectivein the real world, LfD-capable robots should also be able to remember multiplemotion skills. Existing stable-LfD approaches lack the capability ofmulti-skill retention. Although recent work on continual-LfD has shown thathypernetwork-generated neural ordinary differential equation solvers (NODE) canlearn multiple LfD tasks sequentially, this approach lacks stabilityguarantees. We propose an approach for stable continual-LfD in which ahypernetwork generates two networks: a trajectory learning dynamics model, anda trajectory stabilizing Lyapunov function. The introduction of stabilitygenerates convergent trajectories, but more importantly it also greatlyimproves continual learning performance, especially in the size-efficientchunked hypernetworks. With our approach, a single hypernetwork learns stabletrajectories of the robot's end-effector position and orientationsimultaneously, and does so continually for a sequence of real-world LfD taskswithout retraining on past demonstrations. We also propose stochastichypernetwork regularization with a single randomly sampled regularization term,which reduces the cumulative training time cost for N tasks from O$(N^2)$ toO$(N)$ without any loss in performance on real-world tasks. We empiricallyevaluate our approach on the popular LASA dataset, on high-dimensionalextensions of LASA (including up to 32 dimensions) to assess scalability, andon a novel extended robotic task dataset (RoboTasks9) to assess real-worldperformance. In trajectory error metrics, stability metrics and continuallearning metrics our approach performs favorably, compared to other baselines.Our open-source code and datasets are available athttps://github.com/sayantanauddy/clfd-snode.