Keypoints into the Future: Self-Supervised Correspondence in Model-Based Reinforcement Learning

Abstract

Predictive models have been at the core of many robotic systems, from quadrotors to walking robots. However, it has been challenging to develop and apply such models to practical robotic manipulation due to high-dimensional sensory observations such as images. Previous approaches to learning models in the context of robotic manipulation have either learned whole image dynamics or used autoencoders to learn dynamics in a low-dimensional latent state. In this work, we introduce model-based prediction with self-supervised visual correspondence learning, and show that not only is this indeed possible, but demonstrate that these types of predictive models show compelling performance improvements over alternative methods for vision-based RL with autoencoder-type vision training. Through simulation experiments, we demonstrate that our models provide better generalization precision, particularly in 3D scenes, scenes involving occlusion, and in category-generalization. Additionally, we validate that our method effectively transfers to the real world through hardware experiments. Videos and supplementary materials available at https://sites.google.com/view/keypointsintothefuture

References in corpus (7)

• Adam: A Method for Stochastic Optimization
• Visual Foresight: Model-Based Deep Reinforcement Learning for Vision-Based Robotic Control
• Self-Supervised Visual Planning with Temporal Skip Connections
• Learning Predictive Representations for Deformable Objects Using Contrastive Estimation
• Model Predictive Path Integral Control using Covariance Variable Importance Sampling
• A Data-Efficient Approach to Precise and Controlled Pushing
• Experience-Embedded Visual Foresight

Cited by in corpus (5)

• Vision-driven Compliant Manipulation for Reliable, High-Precision Assembly Tasks
• Causal Discovery in Physical Systems from Videos
• 3D Neural Scene Representations for Visuomotor Control
• KeypointDeformer: Unsupervised 3D Keypoint Discovery for Shape Control
• Learning of feature points without additional supervision improves reinforcement learning from images