Planning

How to train RL agents safely? We propose to pretrain a model-based agent in a mix of sandbox environments, then plan pessimistically when finetuning in the target environment.

To plan towards long-term goals through visual prediction, we propose a model based on two key ideas: (i) predict in a goal-conditioned way to restrict planning only to useful sequences, and (ii) recursively decompose the goal-conditioned prediction task into an increasingly fine series of subgoals.

We design and demonstrate a new tactile sensor for in-hand tactile manipulation in a robotic hand.

High-resolution tactile sensing together with visual approaches to prediction and planning with deep neural networks enables high-precision tactile servoing tasks.

In visual prediction tasks, letting your predictive model choose which times to predict does two things: (i) improves prediction quality, and (ii) leads to semantically coherent “bottleneck state” predictions, which are useful for planning.

By exploiting high precision tactile sensing with deep learning, robots can effectively iteratively adjust their grasp configurations to boost grasping performance from 65% to 94%.