Source-linked AI summary
A survey of robot learning from demonstrations for Human-Robot Collaboration
Jangwon Lee
TL;DR
LfD addresses the challenge of teaching robots new tasks without requiring robotics expertise, while collaborative settings add human-centric and communication demands. This survey synthesizes LfD research for human-robot collaboration across communication, interactive learning, and complex-task learning. It concludes that bidirectional communication and human-aware learning are central themes, while noting practical limits in expressing robot intentions and sustaining online teaching.
Problem
Human-robot collaborative LfD must address safety, human feelings and intentions, communication uncertainty, and the added difficulty of learning complex tasks.
Method
The paper surveys recent LfD research for human-robot collaboration and organizes it around communication, Interactive/Active Learning, and complex-task learning.
Results
The survey categorizes communication research into human-to-robot and robot-to-human directions and reviews approaches using social cues, readable behaviors, intention prediction, and language.
Takeaways & Limitations
Bidirectional communication is a central concern for collaborative LfD, alongside treating robots as active partners and addressing human-centric factors.
Takeaways & Limitations
Interactive/Active Learning commonly assumes that human teachers are co-located for online teaching, which can be boring and time consuming.
Abstract
from arXiv · showhide
Robot learning from demonstration (LfD) is a research paradigm that can play an important role in addressing the issue of scaling up robot learning. Since this type of approach enables non-robotics experts can teach robots new knowledge without any professional background of mechanical engineering or computer programming skills, robots can appear in the real world even if it does not have any prior knowledge for any tasks like a new born baby. There is a growing body of literature that employ LfD approach for training robots. In this paper, I present a survey of recent research in this area while focusing on studies for human-robot collaborative tasks. Since there are different aspects between stand-alone tasks and collaborative tasks, researchers should consider these differences to design collaborative robots for more effective and natural human-robot collaboration (HRC). In this regard, many researchers have shown an increased interest in to make better communication framework between robots and humans because communication is a key issue to apply LfD paradigm for human-robot collaboration. I thus review some recent works that focus on designing better communication channels/methods at the first, then deal with another interesting research method, Interactive/Active learning, after that I finally present other recent approaches tackle a more challenging problem, learning of complex tasks, in the last of the paper.
I. INTRODUCTION
LfD enables non-experts to teach robots new tasks, supporting broader deployment beyond research laboratories. This survey focuses on the additional human-centric and communication issues that arise in human-robot collaboration.
- Motivation: LfD can transfer robot prototypes into real-world use because end-users typically need no expert robotics knowledge to teach new tasks.The approach lets users teach robots according to their own requirements.
- Motivation: LfD attracts HRI researchers because robots can learn personal requirements that are difficult to encode through pre-programming.Interaction can also let users correct robot behavior and highlight important task points.
- Human-Robot Collaboration: Collaborative LfD must address human partners’ feelings and intentions in addition to robot learning algorithms and techniques.The survey distinguishes HRI’s communication focus from robotics research focused on reproducing demonstrated arm trajectories.
- Scope: LfD spans machine learning methods and human factors, including the Correspondence Problem and user-friendly demonstration interfaces.The survey concentrates on human-factor issues specific to collaborative tasks rather than common theoretical issues.
- Organization: The survey reviews communication, Interactive/Active Learning, and approaches for teaching complex tasks in successive sections.These topics organize the paper’s treatment of LfD for human-robot collaboration.
II. LFD FOR HUMAN-ROBOT COLLABORATIVE TASKS
LfD for collaboration must address safety, human mental states, and uncertainty arising from unpredictable human and robot behavior. The survey highlights communication, interactive learning, and complex-task learning as central directions.
- Human-Centric Issues: Collaborative robots must address safety, human partners’ mental states, and differing perceptions shaped by relationships and robot appearance.These factors affect whether people feel comfortable working with robot co-workers.
- Human-Centric Issues: Human-centric collaboration problems can be viewed broadly as uncertainty minimization because unpredictable human and robot behaviors create difficulty.The survey identifies reducing communication uncertainty as a key concept for successful learning.
- Communication: HRI research often prioritizes helping users understand robot partners, thereby increasing predictability, rather than only transferring motion knowledge.Understanding intentions can make interactions more comfortable.
- Interactive/Active Learning: Interactive/Active Learning makes teaching bidirectional by having robots provide feedback to human teachers during collaboration.The feedback is used to reduce uncertainty and make learning more efficient.
- Complex Tasks: Learning complex tasks remains a major challenge because the goal is to acquire high-level skills without specifying every arm trajectory.Few studies had investigated complex-task learning specifically for collaborative tasks.
A. Communication
Communication research in collaborative LfD addresses both human-to-robot intention understanding and robot-to-human intention readability. The reviewed approaches use social cues, language, motion prediction, embodied signals, and transparency mechanisms.
- Communication: Collaborative robot communication is bidirectional, but the survey emphasizes research covering both human-to-robot and robot-to-human directions.The paper presents communication as central to robots working around humans.
- User Intention Recognition: Human intention recognition uses social cues such as posture, facial expression, gaze, and verbal signals to guide task learning and movement correction.These cues can reduce the search space for learning new tasks.
- User Intention Recognition: Generalized Grounding Graphs map natural-language commands to robot-relevant groundings such as target objects, places, and paths.After training on annotated demonstrations, the system generated control plans for commands such as placing a tire pallet on a truck.
- User Intention Recognition: A CRF-based energy function addresses ambiguity in natural-language manipulation instructions by encoding language, environment, and controller information.VEIL-300 contains commands, environment information, and ground-truth instruction sequences for six service-robot tasks.
- Robot-to-Human Communication: Embodied cues, including eye movements, can help workers understand a robot’s status and next actions, while perceptions also vary with robot appearance and workplace role.Pre-programmed eye movements following arm trajectories made robot behavior more understandable to industrial workers.
- User Intention Recognition: Early motion prediction combines GMM and GMR with predicted human workspace occupancy to generate safer, more efficient robot motions near people.The approach was demonstrated in PR2 simulation after training on table-top manipulation demonstrations.
- Robot-to-Human Communication: Readable robot behavior conveys intentions, but predictable and legible motions can have contradictory properties because they support different observer inferences.Legibility concerns communicating the robot’s goal, whereas predictability concerns matching observers’ expected behavior.
- Robot-to-Human Communication: Visualization-based transparency mechanisms can help users communicate with robots and understand their abilities, although some users still prefer human-like transparency.The study evaluated the visualization module with 20 participants.
B. Interactive/Active Learning
Interactive/Active Learning treats robot learning as a bidirectional collaboration in which robots reduce uncertainty by querying, requesting help, sensing actively, or training jointly with humans.
- Interactive/Active Learning: Active Learning lets robots ask label, demonstration, or feature queries when uncertain during learning.Participants perceived feature queries as making the robot smartest, and reported them as the most commonly used query type in the study.
- Interactive/Active Learning: A natural-language help system translated detected failures into symbolic requests and then generated assistance requests for a human partner.In table assembly, participants inferred the robot’s requested action better than with general or template-based requests.
- Interactive/Active Learning: Human feedback can train interactive navigation behaviors, but transparency mismatches between observed and actual robot states caused post-training movement problems.The TAMER-based study used positive or negative button feedback to teach five navigation behaviors to Nexi.
- Interactive/Active Learning: Interactive perception combines visual observations with manipulation outcomes to reduce uncertainty over articulated-object models and select informative actions.Experiments covered rotational cabinet doors, a whiteboard eraser, a locked drawer, and stapler grasping, using KL-divergence-based information gain.
- Interactive/Active Learning: Cross-training improves human-robot teamwork by having partners switch roles during collaborative-task training.The place-and-drill study reported significantly improved team performance and more positive robot perceptions than standard reinforcement-learning training.
C. Learning of Complex Tasks
Learning complex collaborative tasks remains challenging, so researchers use skill decomposition, automatic segmentation, and interaction-pattern models to learn structured behavior from demonstrations.
- C. Learning of Complex Tasks: Only a few LfD studies address complex human-robot collaborative tasks, and most decompose them into easier subtasks.This decomposition is presented as the prevailing strategy for making complex tasks learnable.
- C. Learning of Complex Tasks: BP-AR-HMM segments unstructured demonstrations into reusable sub-skills within an integrated framework for complex-task learning.The approach addresses requirements including recognizing repeated skills and generalizing them.
- C. Learning of Complex Tasks: BP-AR-HMM combined with Dynamic Movement Primitives helped a robot learn a multi-step task from unstructured demonstrations.Dynamic Movement Primitives provide a framework for representing dynamical systems.
- C. Learning of Complex Tasks: An extended BP-HMM extracted action primitives and transition probabilities for teaching the sequential task of pizza-dough rolling from human demonstrations.The identified primitives included reach, roll, and reach back.
- C. Learning of Complex Tasks: Applying automatic segmentation to HRC requires extracting interaction-related primitives rather than primitives describing isolated actions.The reviewed non-collaborative approaches illustrate challenges that remain when transferring complex-task learning to collaboration.
- C. Learning of Complex Tasks: Mixtures of Interaction Primitives model multiple nonlinear interaction patterns between human and robot movements from unlabeled collaborative-task demonstrations.Trajectories were represented through weight vectors that parameterized the demonstrated movements.
III. CONCLUSION
The survey identifies communication, interactive learning, and complex-task learning as central themes in LfD for Human-Robot Collaboration. It concludes that collaborative robots must address human-centric issues, robot-specific adaptation, and the difficulty of learning complex tasks.
- III. CONCLUSION: LfD enables non-robotics experts to teach collaborative robots new tasks, but collaboration requires attention to safety, feelings, and intentions.Unlike stand-alone manufacturing tasks, collaborative tasks require robots to work side by side with people.
- III. CONCLUSION: Communication is central because human-robot communication differs from human-human communication and is bidirectional.The survey categorizes communication-focused LfD research into two lines of work.
- III. CONCLUSION: Non-verbal cues and predictable, legible robot motions can convey information and robot intentions during collaboration.Examples include facial expressions, gaze directions, body gestures, and human-readable robot behaviors.
- III. CONCLUSION: Interactive/active learning treats robots as active learners that ask questions or move to gather information when uncertain.This approach uses the robot’s own capabilities during teaching and learning.
- III. CONCLUSION: Learning complex collaborative tasks remains open because robots must discover, generalize, and learn multiple subtasks from demonstrations.HMM-based approaches show possibilities, but existing research generally considers relatively simple tasks.
- III. CONCLUSION: Deep learning has limited application to collaborative LfD because large-scale datasets are difficult to build across different robots, body configurations, and teaching goals.The survey also notes that learned knowledge must be adapted to each robot’s unique form rather than simply imitating humans.