A Survey of Behavior Learning Applications in Robotics -- State of the Art and Perspectives

TL;DR

This survey addresses the state of behavior learning in robotics, focusing on real-world applications for kinematically and sensorily complex robots. It classifies learning problems into manipulation, locomotion, and other behaviors, detailing how learning-based approaches—often RL or imitation-based—are applied across diverse tasks (e.g., peg-in-a-hole, grasping, walking, juggling, and human-robot interaction) and discusses the balance between perception and action, as well as deliberative versus reactive control. The authors highlight key lessons: end-to-end learning shows promise for perception-to-action, but data efficiency, safety, and generalization remain major hurdles; hybrid strategies that integrate classical planning and physics-based priors often yield more robust results. They advocate for knowledge transfer, bootstrapping, and standardized benchmarks to improve comparability and reproducibility, and they argue for a pragmatic blend of learning with traditional robotics to build safer, more capable autonomous systems. The paper concludes that while substantial progress has been made, there is still a long path toward human-level versatility, with future work likely to emphasize sample-efficient learning, high-fidelity perception, and integrated, multi-task systems.

Abstract

Recent success of machine learning in many domains has been overwhelming, which often leads to false expectations regarding the capabilities of behavior learning in robotics. In this survey, we analyze the current state of machine learning for robotic behaviors. We will give a broad overview of behaviors that have been learned and used on real robots. Our focus is on kinematically or sensorially complex robots. That includes humanoid robots or parts of humanoid robots, for example, legged robots or robotic arms. We will classify presented behaviors according to various categories and we will draw conclusions about what can be learned and what should be learned. Furthermore, we will give an outlook on problems that are challenging today but might be solved by machine learning in the future and argue that classical robotics and other approaches from artificial intelligence should be integrated more with machine learning to form complete, autonomous systems.

Figures (8)

• Figure 1: Perception and action. The red background indicates which parts of the behavior are learned. Sometimes both, perception and action, are learned and sometimes only some aspects are learned. The height of each bar indicates complexity of the corresponding part.
• Figure 2: Histogram of publication years of the considered works.
• Figure 3: Mindmap of behavior learning applications. Applications are ordered by domain. Some behaviors are assigned to multiple domains and most of the elementary behaviors could also belong to multiple domains.
• Figure 4: Categorization of manipulation behaviors. Manipulation behaviors are categorized in two dimensions: softness and movability of the manipulated object and dynamics of the behavior. Blue letters indicate the corresponding subsections.
• Figure 5: Learning grasping from sensory information. Exemplary sensor data that could be used to generate grasping behaviors and possible outputs of a skill.