Deep Learning and Machine Learning in Robotics [From the Guest Editors]

Fabio Bonsignorio, David Hsu, Matthew Johnson-Roberson, Jens KoberIEEE Robotics Autom. Mag 2020 · DOI · OpenReview

Introduction

The rapid advancement of deep learning (DL) and machine learning (ML) has fundamentally transformed digital domains such as computer vision and natural language processing. However, transitioning these successes to the physical world of robotics presents unique and formidable challenges. This article, based on a guest editorial for the IEEE Robotics & Automation Magazine, explores the friction between modern ML and physical systems, highlighting curated research efforts that aim to bridge the "sim-to-real" gap.

The Robotics-ML Friction

In virtual environments, data is abundant and errors are inconsequential. In robotics, the paradigm shifts dramatically, creating a fundamental tension between standard ML practices and physical requirements:

  • Sample Efficiency and Hardware Constraints: Unlike simulations that can run millions of iterations in seconds, physical robots are limited by time and mechanical wear. Robots cannot perform trials as easily as software agents.
  • The Labeling Bottleneck: While robots generate vast amounts of raw sensor data, annotating this data (e.g., determining if a grasp was successful) is labor-intensive and expensive.
  • Safety and Reliability: In high-stakes applications like autonomous driving or robotic surgery, a model "hallucination" can lead to catastrophic failure. Safety must be a primary constraint, not an afterthought.
  • Interpretability: The "black box" nature of many DL models poses a barrier to human-robot collaboration. For humans to work effectively alongside robots, the machine's decision-making process must be transparent.

Methodological Innovations

To address these challenges, researchers are developing specialized architectures that incorporate physical and mathematical "priors" to improve efficiency and safety:

  • Geometric Learning: By utilizing Riemannian geometry to represent robot motions, researchers can respect the underlying mathematical manifold of the data, significantly reducing the computational load of learning.
  • Multifidelity Reinforcement Learning: This strategy leverages data from "less realistic" (and thus cheaper) simulations to minimize the amount of dangerous real-world data needed for training.
  • Formal Verification: Integrating pretrained neural networks with reachability analysis ensures that a robot stays within safe operating bounds during runtime, providing a layer of assured monitoring.
  • Interactive and Cross-Modal Learning: New systems allow robots to learn temporal features through direct interaction with human teachers or process complex inputs from optical tactile sensors for 3D exploration.

Real-World Applications

The efficacy of these methods is demonstrated across a broad spectrum of robotic tasks, moving beyond simple laboratory settings:

  • Autonomous Navigation: A deep learning system achieved high-accuracy localization at the Hong Kong International Airport without requiring any modifications to the physical environment.
  • Advanced Manipulation: Multifingered grasp planning has been improved through voxel representations, allowing models to generalize to novel, unseen objects.
  • Medical Rehabilitation: ML-driven "active gravity compensation" is bringing mechatronic rehabilitation systems closer to the performance levels of human physiotherapists.

Conclusion and Future Directions

While significant progress has been made in making robotics more efficient and safe, hurdles remain in interpretability and hardware resilience. Future research is poised to tackle even more extreme frontiers, such as planetary exploration, where the stakes for autonomous decision-making are at their highest. By incorporating structure into learning and prioritizing shared supervision between humans and machines, the field is moving toward a future where DL and ML are as reliable in the physical world as they are in the digital one.

Summary updated

Abstract

Deep learning has gone through massive growth in recent years. In many fields—computer vision, speech recognition, machine translation, game playing, and others—deep learning has brought unprecedented progress and become the method of choice. Will the same happen in robotics and automation? In a sense, it is already happening. Today, deep learning is often the most common keyword for work presented at major robotics conferences. At the same time, robots, as physical systems, pose unique challenges for deep learning in terms of sample efficiency and safety in real-world robot applications. With robots, data are abundant, but labels are sparse and expensive to acquire. Reinforcement learning in principle does not require data labeling but does require a significant number of iterations on real robots. Transferring the capabilities learned in simulation to real robots and collecting sufficient data for practical robot applications both present major challenges. Further, mistakes by robot learning systems are often much more costly than those by their counterparts in the virtual world. These mistakes may cause irreversible damage to robot hardware or, even worse, loss of human lives. Safety is thus paramount for robot learning systems.
BibTeX
@article{bonsignorio2020deep,
  author = {Fabio Bonsignorio and David Hsu and Matthew Johnson-Roberson and Jens Kober},
  title = {{Deep Learning and Machine Learning in Robotics [From the Guest Editors]}},
  journal = {IEEE Robotics Autom. Mag.},
  year = {2020},
  doi = {10.1109/mra.2020.2984470},
}
Figure