Gaussian process implicit surfaces for shape estimation and grasping

Abstract

Abstract-The choice of an adequate object shape representation is critical for efficient grasping and robot manipulation. A good representation has to account for two requirements: it should allow uncertain sensory fusion in a probabilistic way and it should serve as a basis for efficient grasp and motion generation. We consider Gaussian process implicit surface potentials as object shape representations. Sensory observations condition the Gaussian process such that its posterior mean defines an implicit surface which becomes an estimate of the object shape. Uncertain visual, haptic and laser data can equally be fused in the same Gaussian process shape estimate. The resulting implicit surface potential can then be used directly as a basis for a reach and grasp controller, serving as an attractor for the grasp end-effectors and steering the orientation of contact points. Our proposed controller results in a smooth reach and grasp trajectory without strict separation of phases. We validate the shape estimation using Gaussian processes in a simulation on randomly sampled shapes and the grasp controller on a real robot with 7DoF arm and 7DoF hand.

Laser-Radar Data Fusion with Gaussian Process Implicit Surfaces

“…The approach is similar to the one in [8]. The input data of GPIS LR is composed of all training points from each sensing modalities put together:…”

“…However, this approach uses a representation that only allows for a single elevation value at a given (x, y) location, which is not appropriate for representing objects, as mentioned above. A similar fusion approach was used in [8] within a GPIS framework. Raw data from lasers and tactile sensors are directly fused in a GPIS with multi-variance noise in the input dimensions.…”

“…A framework to estimate IS surfaces with uncertainties using Gaussian Process, named GPIS, has been proposed in the literature. This approach has been applied to range sensor data for robotics applications in different contexts, such as change detection [14], active learning [6], and grasping [8].…”

“…We then evaluate and compare the performance of two different fusion approaches within the GPIS framework. The first approach [8], directly fuses raw data from the two different sensor modalities in the GPIS. In the second approach, we fuse points extracted from two initial estimates of the object surface that were built using raw laser data only and raw radar data only, respectively.…”

Dexterous grasping under shape uncertainty

“…Note that ψ is the only parameter in the kernel function and it can be easily determined from the training inputs without any optimization procedure involved, such as maximizing the likelihood. This is one of the main reasons for adopting it here in comparison to other kernel functions such as Radial Basis Function [53,20]. Given the computed covariance matrix and a new data point x * ∈ R 3 , we can use GP to predict the function mean value E(y * ) ∈ R 4×1 and its corresponding variance cov(y * ) ∈ R 4×4 , [50] by…”