Tactile SLAM: Real-time inference of shape
and pose from planar pushing [ICRA 2021]


Sudharshan Suresh1 Maria Bauza2 Kuan-Ting Yu3
Joshua G. Mangelson4 Alberto Rodriguez2 Michael Kaess1
1CMU   2MIT   3XYZ Robotics   4BYU

Tactile SLAM paper
Our method builds a shape contour in real-time and optimizes for pose via geometry and physics-based constraints. Pose-optimization is performed on a factor graph (GTSAM + iSAM2) with fixed-lag smoothing.
The Gaussian process implicit surface represents the zero-level set of the GP potential function. Spatial partitioning enables efficient real-time regression.

Tactile perception is central to robot manipulation in unstructured environments: knowledge of object shape and pose determines the success of generated grasps or nonprehensile actions. Pure tactile perception is challenging: ( i ) touch cannot directly provide global estimates of object shape or pose, ( ii ) the act of sensing itself constantly perturbs the object. We present a method to estimate both object shape and pose in real-time from a stream of tactile measurements. This is applied towards exploration of an unknown object by planar pushing. Our formulation of tactile inference alternates between Gaussian process implicit surface regression and pose estimation on a factor graph. We demonstrate the approach to be real-time, and evaluate on both simulated and real-world planar pushing tasks.

Results


Tactile exploration is simulated in PyBullet on planar objects from the MCube Push Dataset. We use a two-finger pusher for contour following, and collect the tactile measurements and ground-truth poses. The GPIS reconstructs object shape, while geometry and physics constraints inform pose.


We carry out an identical set of exploration tasks with a pusher-slider setup with a single-pusher on an ABB IRB 120. This method can potentially accommodate tactile arrays and vision, and be extended beyond planar pushing.

Example of planar pushing with F/T sensing on plywood surface.

Surface mapping


Contact points condition the GP on zero SDF observations, while contact normals provide function gradient observations. This jointly models both SDF and suface direction for objects. We use the thin-plate kernel, with local GPs for regression.

GP potential function and its implicit surface for noisy contact measurements on the butter shape. The colormap shows spatial grid uncertainty.
The object's extent is split into local GPs with overlapping domains.

Video


Bibtex


            @inproceedings{Suresh21tactile,
              title={Tactile SLAM: Real-time inference of shape and pose from planar pushing}, 
              author={S. Suresh, M. Bauza, K.-T. Yu, J. Mangelson, A. Rodriguez, and M. Kaess},
              booktitle = {Proc. IEEE Intl. Conf. on Robotics and Automation, ICRA},
              month = may,
              year={2021},
            }