[Paper Review] Self-Monitoring Navigation Agent via Auxiliary Progress Estimation

Chih‐Yao Ma, Jiasen Lu+5 more|arXiv (Cornell University)|Jan 10, 2019

Robotic Path Planning Algorithms51 references134 citations

TL;DR

The paper introduces a self-monitoring Vision-and-Language Navigation agent with visual-textual co-grounding and a progress monitor, achieving state-of-the-art results on Room-to-Room, especially unseen environments with an 8% absolute gain in success rate.

ABSTRACT

The Vision-and-Language Navigation (VLN) task entails an agent following navigational instruction in photo-realistic unknown environments. This challenging task demands that the agent be aware of which instruction was completed, which instruction is needed next, which way to go, and its navigation progress towards the goal. In this paper, we introduce a self-monitoring agent with two complementary components: (1) visual-textual co-grounding module to locate the instruction completed in the past, the instruction required for the next action, and the next moving direction from surrounding images and (2) progress monitor to ensure the grounded instruction correctly reflects the navigation progress. We test our self-monitoring agent on a standard benchmark and analyze our proposed approach through a series of ablation studies that elucidate the contributions of the primary components. Using our proposed method, we set the new state of the art by a significant margin (8% absolute increase in success rate on the unseen test set). Code is available at https://github.com/chihyaoma/selfmonitoring-agent .

Motivation & Objective

• Motivate and address VLN where the agent must know which instruction is completed and what is next without explicit goal maps.
• Develop a visual-textual co-grounding module to ground past/upcoming instructions and current motion from surrounding images.
• Introduce a progress monitor to regularize grounding by estimating instruction-following completeness and progress toward the goal.
• Integrate grounding and progress signals into action selection and beam-search inference to improve navigation performance.

Proposed method

• Propose a two-component agent: visual-textual co-grounding and a progress monitor, enabling simultaneous grounding of visuals and instructions and progress estimation.
• Use a sequence-to-sequence LSTM-based architecture with attention to compute grounded textual and visual features at each step.
• Compute textual grounding via soft attention over instruction words with position encoding; ground visual information via attention over panoramic view features.
• Combine grounded instruction and visual context to select actions via inner-product scoring and softmax over navigable directions.
• Introduce a progress monitor that computes a progress signal p_t^{pm} from history, grounded visuals, and textual attention to regularize learning.
• Train with a joint loss combining action-selection cross-entropy and a regression term for progress estimation; inference uses beam search integrating progress signals into beam scoring.

Experimental results

Research questions

• RQ1 How can grounding be performed jointly on visual and textual modalities to determine which instruction is completed and which is needed next?
• RQ2 Can a progress estimation module regularize grounding and improve navigation progress toward the goal in VLN tasks?
• RQ3 Does integrating progress signals into beam search improve unseen-environment generalization in VLN?
• RQ4 What are the contributions of co-grounding and progress monitoring to state-of-the-art performance on R2R?
• RQ5 How data efficiency compare against prior arts with and without augmentation?

Key findings

• Achieves state-of-the-art results on seen and unseen R2R splits, with 8% absolute SR improvement on the unseen test set.
• The co-grounding framework (visual and textual) outperforms the baseline substantially by leveraging shared hidden states for both modalities.
• Progress monitor regularization improves SR on both seen and unseen environments and is crucial for surpassing prior arts without data augmentation.
• Beam-search integration with progress estimation yields additional gains over prior beam-search baselines, particularly for unseen environments.
• Textually grounded attention demonstrates a diagonal-like progression of instruction focus over time, indicating effective grounding of instruction to actions.