Although robotics experts have introduced increasingly complex robotic systems over the past few decades, most solutions introduced to date have been pre-programmed and trained to solve specific tasks. The ability to continuously teach robots new skills while interacting with them could be very beneficial and could promote their widespread use.
Researchers at Arizona State University (ASU) have recently developed a new computing method that allows users to continuously train robots to perform new tasks through dialogue-based interactions. This method was introduced in a paper published on the arXiv preprint server and was originally used to teach robotic manipulators how to successfully prepare cold sandwiches.
Nakul Gopalan, the paper’s guiding author, told Tech Xplore: “Our goal is to contribute to deploying robots in people’s homes that can learn to cook cold meals.” “We want to understand what behaviors people need for home robots from the user’s perspective.
“This user perspective allows us to use language and dialogue when communicating with robots. Unfortunately, these robots may not know everything, such as how to cook pasta for you.”
The main goal of recent work by Gopalan and his colleagues is to design a method that allows robots to quickly acquire previously unknown skills or behaviors from human agents.
In a previous paper presented at the AAAI Artificial Intelligence Conference, the team focused on teaching robots to complete visual tasks through dialogue-based interactions. Their new research builds on previous efforts and introduces a more comprehensive dialogue-based approach to robot training.
“The scope of our work is to improve the applicability of robots by allowing users to personalize their robots,” Weiwei Gu, co-author of the paper, told Tech Xplore. “Because robots need to complete different tasks for different users, and completing these tasks requires different skills, it is impossible for manufacturers to pre-train robots to have all the skills required for all these scenarios. Therefore, robots need to acquire these skills and task-related knowledge from users.”
To ensure that robots can effectively acquire new skills from users, teams must overcome various challenges. First, they must ensure that human users are involved when teaching the robot, and that the robot conveys any questions or requests additional information in a way that non-expert users can understand.
“Secondly, the robot needs to gain knowledge from several interactions with the user, because the user cannot stay with the robot for an infinite length of time,” Gu said. “Finally, despite gaining new knowledge, robots should not forget any pre-existing knowledge.”
Gopalan, Gu and colleagues Suresh Kondepudi and Lixiao Huang set out to work together to address all these requirements for continuous learning. Their proposed interactive continuous learning system handles these three subtasks through three different components.
“First, a dialogue system based on the Large Language Model ( LLM ) asks the user questions to gain any knowledge it may not have or to continue interacting with people,” Gopalan explains. “But how does a robot know that it doesn’t know something?
“To solve this problem, we trained the second component on the robot skill library and learned how they mapped to language commands. If the requested skill is not close to the language the robot already knows, it will ask for a demonstration.”
The team’s newly developed system also includes a mechanism that allows robots to understand when humans demonstrate how to complete a task. If insufficient demonstrations are provided and they have not reliably acquired skills, this module allows robots to request additional skills.
“We jointly use skill representation and language representation to simulate the robot’s skill knowledge,” Gu said. “When a robot needs to perform a skill, it first estimates whether it has the ability to directly perform the skill by comparing the verbal representation of the skill with the verbal representations of all skills the robot has.
“If the robot is confident that it can perform the skill, it will perform the skill directly. Otherwise, it will require the user to perform the skill in person in front of the robot to demonstrate the skill.”
Essentially, after the robot observes a user completing a particular task, the team’s system determines based on the visual information collected that it already has the skills needed to complete the task.
If the system predicts that the robot has not acquired new skills, the robot will ask the user to use the remote control to trace relevant robot trajectories so that it can add these trajectories to the skill library and complete the same tasks independently in the robot. future.
“We associate representations of these skills with LLM and let the robot express its doubts so that even non-expert users can understand the robot’s requirements and provide assistance accordingly,” Gu said.
The second module of the system is based on a pre-trained and fine-tuned action segmented transformer (ACT) with low-rank adaptation (LoRA). Finally, the team developed a continuous learning module that allows the robot to continuously add new skills to its skill pool.
“After the robot is pre-trained with certain pre-selected skills, most of the weights of the neural network are fixed, and only a small portion of the weights introduced by low-level adaptations are used to learn new skills for the robot,” Gu said. “We found that our algorithm was able to effectively learn new skills without disastrously forgetting any pre-existing skills.”
The researchers evaluated their proposed closed-loop skill-learning system in a series of practical tests and applied it to the Franka FR3 robotic manipulator. The robot interacted with eight human users and gradually learned to handle a simple daily task of making sandwiches.
The system developed by Gu, Gopalan and their colleagues will soon be further improved and tested on a wider range of cooking tasks. Researchers are now working to address the rotation problem they observe and expand the range of meals users can teach robots to cook. They also plan to conduct more experiments involving more human participants.
“The issue of rotation is an interesting issue in natural interactions,” Gu added. “This research question also has strong application significance for interactive home robots.
“In addition to solving this problem, we are also interested in expanding the scale of this work by introducing more different tasks and experimenting with our system with users from real-world demographics.”
There are links in the video below this video. If you are interested, you can open it and have a look.
Thank you for watching this video. If you like it, please subscribe and like it. thank
Original text:https://techxplore.com/news/2024-09-approach-robots-skills-dialogue.html
More information: Weiwei Gu et al., Continuous skills and task learning through dialogue, arXiv (2024). DOI：10.48550/arxiv.2409.03166
Journal information: arXiv
Oil tubing: