Visual dexterity: In-hand reorientation of novel and complex object shapes

Abstract

In-hand object reorientation is necessary for performing many dexterous manipulation tasks, such as tool use in less structured environments, which remain beyond the reach of current robots. Prior works built reorientation systems assuming one or many of the following conditions: reorienting only specific objects with simple shapes, limited range of reorientation, slow or quasi-static manipulation, simulation-only results, the need for specialized and costly sensor suites, and other constraints that make the system infeasible for real-world deployment. We present a general object reorientation controller that does not make these assumptions. It uses readings from a single commodity depth camera to dynamically reorient complex and new object shapes by any rotation in real time, with the median reorientation time being close to 7 seconds. The controller was trained using reinforcement learning in simulation and evaluated in the real world on new object shapes not used for training, including the most challenging scenario of reorienting objects held in the air by a downward-facing hand that must counteract gravity during reorientation. Our hardware platform only used open-source components that cost less than 5000 dollars. Although we demonstrate the ability to overcome assumptions in prior work, there is ample scope for improving absolute performance. For instance, the challenging duck-shaped object not used for training was dropped in 56% of the trials. When it was not dropped, our controller reoriented the object within 0.4 radians (23°) 75% of the time.

An Evaluation of AI Integrated Education in Graduate School of Education Using the CIPP Model

“…RL algorithms have been successfully demonstrated for inhand manipulation [9][10][11][21][22][23] tasks. One key difference between these and our work is that in in-hand manipulation, the object typically starts in close proximity to the robot, whereas in our domain, we must solve the additional problem of reaching and making contact with the object.…”

“…Like the recent approaches for contact-rich tasks such as in-hand manipulation [9][10][11] and locomotion [12,13], we use a simulator to train a policy and then transfer it to the real world. In this scheme, there are two primary challenges that we need to address for non-prehensile manipulation: exploration and the sim-to-real gap.…”

“…The real robot has safety measures that prevent the robot to move above a certain speed and contact force. To account for this, we can terminate the episode if the robot violates them during learning, as in [10]. However, we found this makes learning difficult as it explores too conservatively, and fails to learn a policy in some of our domains.…”

Learning Generalizable Tool Use with Non-Rigid Grasp-Pose Registration

“…Secondly, imitation can be used to transfer knowledge between policies, also known as policy distillation (PD) [18]. A specific formulation of this framework has recently proven useful in the context of robot learning [1], [2], wherein a teacher policy is trained from privileged, low-dimensional observations and then distilled into a visual student policy. The distillation from teacher to student is performed using DAgger [19].…”

“…observations pushes computational and sample complexities beyond currently feasible limits. To train visual policies more efficiently, policy distillation (PD) has recently been used to transfer knowledge from policies trained on low-dimensional states to high-dimensional visual observations [1], [2]. However, differences between the characteristics of these observation spaces and resulting optimal behaviors have not yet been considered.…”