Deep Object-Centric Representations for Generalizable Robot Learning

Abstract

Robotic manipulation in complex open-world scenarios requires both reliable physical manipulation skills and effective and generalizable perception. In this paper, we propose a method where general purpose pretrained visual models serve as an object-centric prior for the perception system of a learned policy. We devise an object-level attentional mechanism that can be used to determine relevant objects from a few trajectories or demonstrations, and then immediately incorporate those objects into a learned policy. A task-independent meta-attention locates possible objects in the scene, and a task-specific attention identifies which objects are predictive of the trajectories. The scope of the task-specific attention is easily adjusted by showing demonstrations with distractor objects or with diverse relevant objects. Our results indicate that this approach exhibits good generalization across object instances using very few samples, and can be used to learn a variety of manipulation tasks using reinforcement learning.

Cleaning Tasks Knowledge Transfer Between Heterogeneous Robots: a Deep Learning Approach

“…As a consequence, the robot can become stuck performing the same trajectory over and over again when the performed movement does not significantly change the dirt configuration: this typically happens when the table is almost cleaned, as discussed in the previous section Implementing a stopping criteria based on the detection of a clean table could be a straightforward solution for this problem. Moreover, a direction for further research that can alleviate this issue is to use a deep reinforcement learning approach [8], where the robot can learn from trial and error how to clean the table, based on a set of image features provided by a deep neural network.…”

“…The EM training procedure finds, in the E-Step, responsibilities γ n,i = π i p(ξ n |i) K k=1 π k p(ξ n |k) (9) and uses these values to update estimates for parameters Z µ i,j , Z Σ i,j and π i in the M-Step (for more details please refer to [3]). After learning, given a new set of frames of reference X j = {A j , b j }, provided by the neural network from the test image, a trajectory T n is generated in the task space by conditioning the distribution p(ξ|·) on the time variable t n , using (5), (7) and (8).…”

“…Unfortunately, typical cleaning actions, such as wiping, washing, sweeping and scrubbing, require a robot with fine manipulation abilities to be performed [23]. As a result, several researchers focus their attention on service robots equipped with manipulators to perform cleaning tasks [32,42,25,15,31,9,33,41,22,24,21,14,27,7,8,26,35,19,11,10,37,40,1,16,39,18,38].…”

“…In such situations deep reinforcement learning (Deep RL) models can simultaneously learn a desired behaviour from self exploration and extract the relevant features from raw data. Devin et al [8] developed a Deep RL object-level attentional mechanism used to control a robot in different tasks like pouring almonds in a cup or sweeping citrus from a table. Moreover, Liu et al proposed an imitation-from-observation algorithm used to perform various pushing and swiping actions.…”

Untangling Sequences: Behavior vs. External Causes

“…It might in theory be possible to extend those algorithms to model combined sensorimotor sequences. However existing machine learning methods for robot sensorimotor learning (Pinto et al, 2016;Devin et al, 2017) are significantly different from machine learning techniques for temporal sequences (Waibel, 1989;Hochreiter and Schmidhuber, 1997;Fine et al, 1998). A detailed comparison of sequence learning techniques to our model can be found in (Cui et al, 2016;Hawkins and Ahmad, 2016).…”