A System for General In-Hand Object Re-Orientation

Abstract

Figure 1: We present a simple framework for learning policies for reorienting a large number of objects in scenarios where the (1) hand faces upward, (2) hand faces downward with a table below the hand and (3) without the support of the table. The object orientation in the rightmost image in each row shows the target orientation.

Robot Target Location Based on the Difference in Monocular Vision Projection

“…This study introduced QT-Opten and realized closed-loop vision-based control, whereby the robot continuously updates its grasp strategy based on the most recent observations to optimize long-horizon grasp success. Chen et al [19] proposed a simple model-free framework that can learn to redirect objects when the robot hand is upward and downward and has strong zero sample migration performance. Pertsch et al [20] proposed a deep latent variable model, which combines the embedding space of learning skills with the a priori skill from the experience of off-line agents.…”

Rapid Locomotion via Reinforcement Learning

“…Teacher-student training enables the agent to specialize its behavior to the current dynamics d t , instead of learning a single behavior that works across different d t . This so-called implicit system identification approach has been previously developed in a number of works involving object re-orientation with a multi-finger hand [8], self-driving cars [7] and locomotion [23,24,27,29]. Like work applying student-teacher learning to blind walking [23,24], our teacher policy observes d t , the dynamic properties of the robot and terrain.…”

“…In [27], the student policy observed a forward-facing depth image. [8] applied the teacher-student training approach to the task of object reorientation using a dexterous five-fingered hand. In this work, d t included the true position of the object as well as the ground-truth state of the hand's fingers.…”

Multimodal perception for dexterous manipulation

“…Behavioral synthesis. Data-driven approaches have consistently used Reinforcement Learning (RL) on joint-based control to solve complex dexterous manipulation in robotics (Rajeswaran et al, 2018;Kumar et al, 2016;Nagabandi et al, 2019;Chen et al, 2021). In order to yield more naturalistic movements, different methods have leveraged motion capture data (Merel et al, 2017;Hasenclever et al, 2020).…”

“…In addition to MyoDex π # , we also train a baseline agent using π * expert-student method (Jain et al, 2019;Chen et al, 2021). Individual task-specific policies (π i ) were used as experts.…”