In this work we present T-Dex, a new approach for tactile-based dexterity, that operates in two phases. In the first phase, we collect 2.5 hours of play data, which is used to train self-supervised tactile encoders. In the second phase, given a handful of demonstrations for a dexterous task, we learn non-parametric policies that combine the tactile observations with visual ones. Across five challenging dexterous tasks, we observe that our tactile-based dexterity models outperform purely vision and torque-based models by an average of 1.7X.
We evaluate our framework on several dexterous tasks that are difficult to solve with visual information alone.
We observe that our tactile-based policies are able to generalize on new objects that are structurally similar but visually different. Our policy also fails in some cases, which is indicated by a red cross.
Robot rollout for cup unstacking task on unseen cups.
Robot rollout for bowl unstacking task on unseen bowls.
T-Dex consists of two phases: First phase is to collect a tactile play dataset, where a teleoperator collects 2.5 hours of aimless demonstrations. Given this dataset, a tactile encoder is trained using a self-supervised objective. In the second phase a task-specific policy is acquired by leveraging the pre-trained tactile encoder in the first phase.
Examples of play data collected by a human teleoperator.
Examples of tactile readings from highest activated tactile pads.
We observe that we require both the tactile and image observations combined together to successfully complete our tasks. We show the importance of our method decisions by comparing image only and tactile only neighbor matching policies to T-Dex.
Success: It successfully slides the inner cup.
Failure: It fails since the policy doesn't know how much force should be applied.
Failure: It fails since the policy doesn't know where the object is.
@misc{guzey2023dexterity,
title={Dexterity from Touch: Self-Supervised Pre-Training of Tactile Representations with Robotic Play},
author={Irmak Guzey and Ben Evans and Soumith Chintala and Lerrel Pinto},
year={2023},
eprint={2303.12076},
archivePrefix={arXiv},
primaryClass={cs.RO}
}
We show all the robot rollouts and the robot experiments for each of our tasks.