[Paper Review] From Machine Learning to Robotics: Challenges and Opportunities for Embodied Intelligence

Nicholas Roy, Ingmar Posner+18 more|arXiv (Cornell University)|Oct 28, 2021

Reinforcement Learning in RoboticsComputer Science185 references33 citations

TL;DR

The paper argues that embodied intelligence requires distinct learning approaches from traditional ML, outlining key challenges, inductive biases, and research directions to enable robust, safe, and generalizable robot learning. It also suggests leveraging dual-process-inspired architectures and compositional representations to bridge System 1 and System 2 reasoning.

ABSTRACT

Machine learning has long since become a keystone technology, accelerating science and applications in a broad range of domains. Consequently, the notion of applying learning methods to a particular problem set has become an established and valuable modus operandi to advance a particular field. In this article we argue that such an approach does not straightforwardly extended to robotics -- or to embodied intelligence more generally: systems which engage in a purposeful exchange of energy and information with a physical environment. In particular, the purview of embodied intelligent agents extends significantly beyond the typical considerations of main-stream machine learning approaches, which typically (i) do not consider operation under conditions significantly different from those encountered during training; (ii) do not consider the often substantial, long-lasting and potentially safety-critical nature of interactions during learning and deployment; (iii) do not require ready adaptation to novel tasks while at the same time (iv) effectively and efficiently curating and extending their models of the world through targeted and deliberate actions. In reality, therefore, these limitations result in learning-based systems which suffer from many of the same operational shortcomings as more traditional, engineering-based approaches when deployed on a robot outside a well defined, and often narrow operating envelope. Contrary to viewing embodied intelligence as another application domain for machine learning, here we argue that it is in fact a key driver for the advancement of machine learning technology. In this article our goal is to highlight challenges and opportunities that are specific to embodied intelligence and to propose research directions which may significantly advance the state-of-the-art in robot learning.

Motivation & Objective

• Define embodied intelligence as the energy-and-information exchange with a physical environment and motivate why it differs from standard ML settings.
• Identify the core challenges of learning for embodied agents, including safety, non-stationarity, and finite energy constraints.
• Argue for inductive biases and architectural principles that support generalization across tasks and environments.
• Propose research directions including core knowledge, hierarchical abstractions, compositional representations, and morphology-aware learning.
• Discuss evaluation and verification challenges for embodied intelligent systems.

Proposed method

• Review and synthesize perspectives from robotics, machine learning, cognitive science, and related fields to define embodied learning requirements.
• Characterize inductive biases specific to embodied agents and discuss their implications for architecture and learning.
• Introduce the Dual Process Theory-inspired framing (System 1 and System 2) as a blueprint for combining fast, learned policies with deliberative planning.
• Explore the role of composition, causality, and physics-inspired priors as structural inductive biases.
• Advocate for meta-learning, curricula, and data-efficient strategies to cope with non-stationary, real-world environments.

Experimental results

Research questions

• RQ1 What inductive biases enable embodied agents to learn effectively and robustly across changing tasks and environments?
• RQ2 How should architectures for embodied intelligence balance fast, reflex-like responses with slower, deliberative reasoning?
• RQ3 What representations and compositional structures support generalization and data efficiency in robotics?
• RQ4 How does agent morphology influence learning and what are effective methods to integrate morphology into learning pipelines?
• RQ5 How can we evaluate and verify embodied intelligent systems operating in non-stationary real-world settings?

Key findings

• Robust embodied learning requires inductive biases tied to action and perception, not just standard ML priors.
• Dual Process Theory motivates architectures that combine fast, intuitive policies with slower, deliberate inference and planning.
• Compositional representations and causality-aware models are crucial for generalization in non-stationary environments.
• Learning architectures should integrate multi-level abstractions to manage computational demands while preserving task-specific performance.
• Morphology (sensors, actuators, energy constraints) significantly shapes what an agent can learn and do, necessitating integrated design and learning approaches.
• Evaluation and verification of embodied learners remain open challenges requiring principled frameworks.