[Paper Review] On Evaluation of Embodied Navigation Agents

Peter Anderson, Anne Lynn S. Chang+9 more|arXiv (Cornell University)|Jul 18, 2018

Robotic Path Planning AlgorithmsComputer Science28 references503 citations

TL;DR

A consensus paper proposing standardized task formats, evaluation metrics, generalization regimes, and standard benchmarks for embodied navigation in 3D environments, with SPL as the recommended primary metric.

ABSTRACT

Skillful mobile operation in three-dimensional environments is a primary topic of study in Artificial Intelligence. The past two years have seen a surge of creative work on navigation. This creative output has produced a plethora of sometimes incompatible task definitions and evaluation protocols. To coordinate ongoing and future research in this area, we have convened a working group to study empirical methodology in navigation research. The present document summarizes the consensus recommendations of this working group. We discuss different problem statements and the role of generalization, present evaluation measures, and provide standard scenarios that can be used for benchmarking.

Motivation & Objective

• Clarify problem statements and goal types in embodied navigation (PointGoal, ObjectGoal, AreaGoal).
• Propose rigorous generalization and exploration regimes with quantified prior exposure to test environments.
• Recommend a single, interpretable primary evaluation metric (SPL) and supportive auxiliary metrics.
• Advocate for continuous-space simulators, SI-unit reporting, and open-source deployment to real robots.
• Provide standard benchmark scenarios across multiple datasets to enable reproducible comparisons.

Proposed method

• Define three goal types (PointGoal, ObjectGoal, AreaGoal) and discuss specification modalities (coordinates, categories, images, language).
• Outline generalization regimes (no prior exploration, pre-recorded exploration, time-limited exploration) and quantify exposure prior to evaluation.
• Introduce SPL (Success weighted by Inverse Path Length) as the primary navigation metric and specify a DONE action for task completion to ensure understanding of goal achievement.
• Recommend continuous state spaces and SI units in simulators; emphasize open-source tooling to bridge simulation and real robots.
• Provide standard scenarios drawn from SUNCG, Matterport3D, AI2-THOR, and Gibson with train/validation/test splits for reproducible benchmarking.
• Encourage reporting auxiliary metrics alongside SPL for a fuller performance picture.

Experimental results

Research questions

• RQ1 What are robust, common task definitions for embodied navigation that support cross-study comparisons?
• RQ2 How should generalization to new or partially explored environments be quantified and reported?
• RQ3 What is an appropriate, interpretable primary metric to evaluate navigation performance across diverse scenes?
• RQ4 How should simulation platforms be designed to facilitate transfer to real-world robots?
• RQ5 What standard scenarios can support reproducible benchmarking across multiple indoor environments?

Key findings

• SPL is proposed as the primary, interpretable navigation performance metric, with a binary success signal based on a DONE action and geodesic distance to the goal.
• Geodesic distance rather than Euclidean distance should be used to evaluate proximity to the goal, accounting for environment structure.
• A DONE action must be produced to consider an episode successful, ensuring the agent’s understanding of goal completion.
• Simulation-based benchmarks should use continuous state spaces and SI units to improve realism and interoperability; open-source deployment tools are encouraged to facilitate transfer to real robots.
• Standardized scenarios are provided across SUNCG, Matterport3D, AI2-THOR, and Gibson with train/validation/test splits to enable reproducible comparisons.