HyperDynamics: Meta-Learning Object and Agent Dynamics with Hypernetworks
This work addresses the challenge of adapting dynamics models to varying environments in robotics and control, offering a novel approach that improves performance and generalization, though it is incremental in building on meta-learning and hypernetworks.
The authors tackled the problem of learning dynamics models that adapt to environment variations by proposing HyperDynamics, a meta-learning framework that generates neural dynamics parameters based on inferred system properties from interactions and visual observations, and it outperformed existing methods on object pushing and locomotion tasks while matching ensemble performance and generalizing to unseen variations.
We propose HyperDynamics, a dynamics meta-learning framework that conditions on an agent's interactions with the environment and optionally its visual observations, and generates the parameters of neural dynamics models based on inferred properties of the dynamical system. Physical and visual properties of the environment that are not part of the low-dimensional state yet affect its temporal dynamics are inferred from the interaction history and visual observations, and are implicitly captured in the generated parameters. We test HyperDynamics on a set of object pushing and locomotion tasks. It outperforms existing dynamics models in the literature that adapt to environment variations by learning dynamics over high dimensional visual observations, capturing the interactions of the agent in recurrent state representations, or using gradient-based meta-optimization. We also show our method matches the performance of an ensemble of separately trained experts, while also being able to generalize well to unseen environment variations at test time. We attribute its good performance to the multiplicative interactions between the inferred system properties -- captured in the generated parameters -- and the low-dimensional state representation of the dynamical system.