Bhardwaj, Amit

Remotely operated robotic systems have gained importance in executing tasks in complex and challenging environments which are difficult to automate. This paper focuses on developing reference hardware and software architectures for haptic-based teleoperation under various physical and network conditions. The system consists of a 6-DOF haptic device as the leader and a 6-DOF robotic manipulator as the follower. The control architecture used for teleoperation is Jacobian inverse control, which enables the follower to follow the leader when commanded. The performance of the proposed architecture is determined in terms of the error between current and commanded motion for different input velocities, communication delays in different network configurations, and the stable haptic force feedback at the haptic end.

Data-driven modeling and rendering is a general approach in haptics aiming to provide highly accurate haptic perceptual experiences simulating complex real physical dynamics, such as deformable or textured objects. A prevalent problem in the present methods for data-driven haptics is that the computational cost for modeling grows rapidly, even becoming intractable, as the interaction complexity or the number of data increases. This paper proposes one data-driven method featured with greatly improved computational efficiency for modeling viscoelastic deformable objects. This advantage is enabled by the use of fractional derivatives for modeling features and regression forests for data-interpolation models. For the benchmark of normal interaction on deformable objects, we describe a computational framework for data-driven haptic modeling and rendering. Its performance is validated by physical experiments for modeling accuracy and cost and a perceptual experiment for the similarity between real and virtual objects. The experiments demonstrate that our method offers highly realistic haptic perceptual experiences with markedly better modeling cost (at least ten times) than other state-of-the-art methods.

A perceptual adaptive sampling approach is used for haptic data reduction for a typical teleoperation application. The approach is based on the Weber's law of perception and is widely named as the 'perceptual deadband' approach. The approach transmits only those samples which are perceptually significant from the last perceived stimulus. Hence, the approach considers that the perception of the current stimulus depends on the last perceived force. In this work, we question this consideration and hypothesize that the perception of a change in stimulus is a function of the current and previous stimulus (irrespective of whether it is perceived or not). For the purpose, an extensive psychophysical experiment is designed where users are subjected to a stair-case force stimuli and are asked to respond to the change in stimuli. Perceived and non-perceived responses are recorded for several users. A machine learning classifier is trained to predict the label of the responses. Our findings suggest that the perception of the current force stimulus depends on the previous force stimuli, hence it validates our hypothesis that the assumption of the current structure of the perceptual deadband approach for haptic data reduction has limitations.

Haptic rendering of virtual objects requires a model for the object's hardness. The classical spring-damper model is effective in resisting the user's penetration into the virtual surface, while a short vibratory transient signal superimposed at contact increases the user-perceived hardness. In this paper, we are concerned with the combined case: the perceived hardness of a virtual surface rendered by the spring-damper model with a contact transient, in the tool-mediated perception. We report three perceptual experiments conducted to clarify the effects of five rendering parameters - stiffness and damping (spring-damper model); and amplitude, frequency, and decay rate (contact transient). All the parameters except decay rate are shown to have statistically significant effects on the perceived hardness. We also present a psychophysical magnitude function of perceived hardness for the four significant parameters of stiffness, damping, amplitude, and frequency. This mathematical model helps virtual environment designs build the virtual surfaces of desired perceived hardness.

In the literature, perceptual deadband approach based on the Weber's law of perception has been employed to reduce haptic data (i.e., force) rate for a typical teleoperation application. The approach selects only those samples for transmission which lie outside the perceptual deadband. The existing structure of the deadband has linear decision boundaries and assume that the just noticeable differences (JNDs) for increasing and decreasing change in a reference force stimulus are similar. This paper questions this assumption and searches for an asymmetric perceptual deadband using a data-driven approach. For the purpose, we design an experimental setup for collecting human haptic responses (perceived and non-perceived) of several users for a force range [3, 5] N. A machine learning classifier inspired from the Weber's law is trained to predict the labels of the responses and define a generalized linear perceptual deadband for each user. The results show that the generalized deadband do provide different increasing and decreasing JND (i.e., pointing towards asymmetry in force perception), but fails to improve the data reduction significantly as compared to the existing symmetric one.

This paper proposes an alternative data-driven haptic modeling method of homogeneous deformable objects based on a CatBoost approach-a variant of gradient boosting machine learning approach. In this approach, decision trees are trained sequentially to learn the required mapping function for modeling the objects. The model is trained on the input feature vectors consisting of position, velocity and filtered velocity samples to estimate the response force. Our approach is validated with a publicly available two-finger grasping dataset. The proposed approach can model unknown interactions with good accuracy (relative root mean squared error, absolute relative error and maximum error less than 0.06, 0.18 and 0.76 N, respectively) when trained on just 20% of the training data. The CatBoost-based method outperforms the existing data-driven methods both in terms of the prediction accuracy and the modeling time when trained on similar size of the training data.