A cognitive robot equipped with autonomous tool innovation expertise

Abstract

Like a human, a robot may benefit from being able to use a tool to solve 
 a complex task. When an appropriate tool is not available, a very useful ability for a robot is to create a novel one based on its experience.
 With the advent of inexpensive 3D printing, it is now possible to give robots such an ability, at least to create simple tools. We proposed a method for learning how to use an object as a tool and, if needed, to design and construct a new tool. The robot began by learning an action model of tool use for
 a PDDL planner by observing a trainer. It then refined the model by learning by trial and error. Tool creation consisted of generalising an existing tool model and generating a novel tool by instantiating the general model. Further learning by experimentation was performed. Reducing the search space of potentially useful tools could be achieved by providing a tool ontology.
 We then used a constraint solver to obtain numerical parameters from abstract descriptions and use them for a ready-to-print design. We evaluated our system using a simulated and a real Baxter robot in two cases: hook and wedge. We found that our system performs tool creation successfully.

Creative Problem Solving in Artificially Intelligent Agents: A Survey and Framework

“…In contrast to the construction of navigational structures, prior work by Nair et al Nair, 2020) introduce the Robogyver architecture that extends classical planning through supervised learning, to enable a robot to create or "macgyver" novel tools from available objects. In similar work, Wicaksono and Sammut (Wicaksono & Sheh, 2017;Wicaksono & Sammut, 2020) integrate planning within the CREATIVE architecture, to enable a robot to craft novel tools through 3D printing, as opposed to using available objects. Lastly, in the cognitive architecture of SOAR, Lieto et al (Lieto, Perrone, Pozzato, & Chiodino, 2019) show that concept representation in a knowledge base can be used as a means for "subgoal" resolution (or plan repair) within planning.…”

“…Planning and execution is then performed through a modified planner in the ICARUS architecture, which take these quantitative attributes into consideration. Prior work by Wickasono and Sammut (Wicaksono & Sheh, 2017;Wicaksono & Sammut, 2020) capture properties of tools within a hierarchical planning language that is then used to create novel tools. Their hierarchy captures aspects such as length, width, and shape of the tools.…”

“…In Chitnis et al (Chitnis et al, 2021), the agent builds a world transition model through state-based goal exploration, where selected goals are optimized on a novelty measure. In the space of tool creation, prior work by Wickasono and Sammut (Wicaksono & Sheh, 2017;Wicaksono & Sammut, 2020) use heuristics to guide autonomous tool creation by a robot, where the heuristic prioritizes new tools that are similar to an existing one, where the similarity between tool representations is computed as the number of edit operations needed to transform one representation into the other. Nair et al use supervised learning techniques to compute object fitness scores that are then incorporated into existing planning heuristics within A * planning to enable a robot to perform tool construction.…”

“…Several existing works in CPS have focused on evaluating the models on various physical platforms such as Sawyer (Xie et al, 2019;Xu et al, 2018), Baxter (Wicaksono & Sheh, 2017;Wicaksono & Sammut, 2020;Yang et al, 2020), Kinova (Fitzgerald et al, 2017(Fitzgerald et al, , 2019, PR2 (Gajewski et al, 2019;Murooka et al, 2019), KUKA (Hangl et al, 2017), Darias (Kroemer et al, 2015), and custom platforms that are specifically built for the tasks that the robot is required to perform (Saboia da Silva et al, 2019;Boteanu et al, 2015;Tosun et al, 2018;Hangl et al, 2017). The Sawyer and Kinova platforms each have a single 7-DOF robot arm with a stationary base.…”

Visual Commands for Control of Food Assistance Robot

“…The signs are recognized using CNN. A simulation environment is used to validate the algorithm and to prevent accidents with the user, a fairly common way for robot training [20], [21], as well as the adequacy of the robotic end-effector [22], which in this case is designed spoon type. This work helps to expand the state of the art in the development of assistance robots to improve the quality of life and use deep learning techniques for their training.…”