Skip to the content.


We introduce a new method for exploration in robot manipulators. We show that this method works well on sparse-reward tasks as pretraining task.

The paper is available on OpenReview - and an older version on arXiv -

Here’s the general idea in a little illustration.

  1. The robot learns to associate visual stimuli with expected haptic sensations.
  2. The robot is surprised by a haptic sensation feeling different than expected. We use this difference as reward signal for exploration.

method illustration, showing visual surpise from unexpected haptic sensation

Here’s an outline of our architecture.

network diagram


TODO - we have a little pybullet live demo in progress that’ll go here

Minitouch Benchmark

For this task, we designed a new small interaction benchmark that can be found here: MiniTouch -

Paper Code

TODO - the paper code is currently WIP and will be released shortly