Mind the Error! Detection and Localization of Instruction Errors in Vision-and-Language Navigation

Francesco Taioli1,4, Stefano Rosa2, Alberto Castellini1, Lorenzo Natale2,
Alessio Del Bue2, Alessandro Farinelli1, Marco Cristani1, Yiming Wang3
1 University of Verona, 2 Istituto Italiano di Tecnologia (IIT), 3 Fondazione Bruno Kessler, 4 Polytechnic of Turin
Code (to be released) Dataset (to be released)
Teaser

Full architecture of our IEDL method.

Abstract

Vision-and-Language Navigation in Continuous Environments (VLN-CE) is one of the most intuitive yet challenging embodied AI tasks. Agents are tasked to navigate towards a target goal by executing a set of low-level actions, following a series of natural language instructions. All VLN-CE methods in the literature assume that language instructions are exact. However, in practice, instructions given by humans can contain errors when describing a spatial environment due to inaccurate memory or confusion. Current VLN-CE benchmarks do not address this scenario, making the state-of-the-art methods in VLN-CE fragile in the presence of erroneous instructions from human users. For the first time, we propose a novel benchmark dataset that introduces various types of instruction errors considering potential human causes. This benchmark provides valuable insight into the robustness of VLN systems in continuous environments. We observe a noticeable performance drop (up to -25%) in Success Rate when evaluating the state-of-the-art VLN-CE methods on our benchmark. Moreover, we formally define the task of Instruction Error Detection and Localization, and establish an evaluation protocol on top of our benchmark dataset. We also propose an effective method, based on a cross-modal transformer architecture, that achieves the best performance in error detection and localization, compared to baselines. Surprisingly, our proposed method has revealed errors in the validation set of the two commonly used datasets for VLN-CE, i.e., R2R-CE and RxR-CE, demonstrating the utility of our technique in other tasks.

Scenario

Teaser

An agent must follow instructions expressed in natural language to reach a target goal. For example: “Exit the bathroom and go left (✓right), then turn left at the big clock and go into the bedroom and wait next to the bed.” In this case, just changing “right” to “left” causes the agent to stop the exploration in the wrong location, even if during the path it did not see the “big clock” (yellow star).

Effect of Direction error

We show the full episode for the example reported in the Scenario section above (also Fig.1 in the paper).

The videos shows the agent following the perturbed and correct instructions respectively.
Perturbed Instruction
" Exit the bathroom and go left, then turn left at the big clock and go into the bedroom and wait next to the bed."
Correct Instruction
"Exit the bathroom and go left, then turn right at the big clock and go into the bedroom and wait next to the bed."


IEDL for detecting errors on R2R-CE standard Validation set

We answer the following question: What if we apply IEDL on the R2R-CE dataset?

Our method helped to detect 8 episodes from R2R-CE Val Unseen dataset that present different issues.

Here we show the BEVBert trajectories for these episodes, as well as the detected issue.


Note: any typos in the instructions are reported verbatim from the dataset.
Target error
"Walk up to the photo on the wall directly in front of you."
Ep id: 327
Ambiguous Instruction
"Turn around and go through the archway. Turn left and take the extreme left. Stop near the mantle."
Ep id: 1214
Instruction error
"Walk past beige rug. Walk past butler's pantry. Make left after eye chart. Wait at gold polka dot picture frame."
Ep id: 249
Instruction error
"Go through middle room and to the right into the room with the sink, go straight through until you come to an archway on your left make a left through it and go through the doorway and wait."
Ep id: 866

Instruction error
"Turn right to face an old pew and chairs. Walk up to the old chairs and turn right. Walk down the row until you reach the 5th chair."
Ep id:1625
Instruction error & Mesh problem
"Go through the dooorway on the right, continue straihtacross the hallway and into the room ahead. Stop near the shell."
Ep id: 30

Reconstruction problem
"Go to the ottoman. Go to the bed. Go to the wardrobe. Go to the glass jug."
Ep id: 503
Instruction error & Mesh problem
"Turn right and go through the door. Go down the hall way and go up the stairway to the left. Continue up the stairway to the left. Wait by the statue."
Ep id: 825

IEDL for detecting errors on RxR-CE standard Validation set

We report also the episode to be removed from the validation split of RxR-CE.

To prevent slow page loading due to longer episode lengths (resulting in heavier videos) and a higher number of instruction tokens (averaging 110 per episode) in RxR-CE compared to R2R-CE, we display only two videos on the main page.

The remaining videos are accessible here.

Instruction error
"Came out the wash room beside the glass window and curtain also. Go straight in front of mirrors and beside the big watch.Again go straight and turn wright come to another room"
Ep id: 5877
Target Error
"Too long, refer to the video caption"
Ep id: 6935