Lightweight, Uncertainty-Aware Conformalized Visual Odometry

Alex C. Stutts, Danilo Erricolo, Theja Tulabandhula, Amit Ranjan Trivedi

TL;DR

This paper introduces a novel framework for visual odometry (VO) that provides statistically guaranteed uncertainty estimates for robot pose predictions. By leveraging conformal inference, the approach offers lightweight and adaptable uncertainty bands crucial for risk-aware decision-making in resource-constrained edge robotics.

What Problem Does This Paper Address?

Autonomous robots, especially those operating in safety-critical environments like drones or surgical robots, need to not only predict their position and orientation (pose) accurately but also understand the confidence in those predictions. This is known as estimating predictive uncertainty. Traditional methods for uncertainty estimation, like Bayesian neural networks, are often computationally too expensive for edge robotics devices with limited processing power and battery life. Furthermore, existing methods may struggle to provide reliable uncertainty estimates, especially when dealing with real-world data distortions (aleatoric uncertainty) or when the model's knowledge is limited (epistemic uncertainty). This paper aims to bridge this gap by developing a computationally efficient and statistically robust framework for uncertainty-aware visual odometry suitable for these resource-constrained platforms.

Methods

The core of this paper lies in applying conformal inference (CI) to visual odometry. CI is a distribution-free approach that provides statistically valid uncertainty intervals without making strong assumptions about the underlying data distribution. The authors propose four novel methodologies:

  • Univariate Conformalized Quantile Regression (CQR): This method applies CQR to each dimension of the pose (position and orientation) independently. It then combines these univariate uncertainty bands to form a rectangular uncertainty region.
  • Conformalized Set Prediction (CSP): This approach reframes VO as a classification problem. By discretizing the pose space and using softmax outputs from a neural network, it generates prediction sets that can be disjoint, offering a different type of uncertainty representation.
  • Multivariate CQR with MC-Dropout (MCQR w/ MC-dropout): This method extends CQR to handle correlations between pose dimensions by using a conditional variational autoencoder (CVAE) and incorporating Monte Carlo dropout (MC-dropout) for data augmentation and improved uncertainty distillation.
  • Conformalized Joint Prediction (CJP): This is the most advanced method, which jointly trains the pose prediction and uncertainty band estimation using a novel loss function. This loss function combines traditional MSE and KL divergence with interval scoring and a combined calibration loss to achieve a balance between prediction interval calibration and sharpness. This approach is designed for efficiency, as it shares network layers for both tasks.

Datasets

The paper mentions using Scene 2 of the RGB-D Scenes Dataset v.2 for evaluating the proposed methods. They also refer to general VO datasets and the need for a calibration set, which is split from the training data.

Key Insights and Ideas

  • Conformal Inference is a powerful tool for uncertainty quantification in robotics: CI offers statistically guaranteed uncertainty intervals without requiring complex Bayesian models, making it ideal for edge robotics.
  • Univariate methods are simple but can be overly conservative: While easy to implement, treating each pose dimension independently can lead to unnecessarily wide uncertainty bands.
  • Multivariate methods capture inter-dimensional correlations: Accounting for how different pose components relate to each other leads to more informative uncertainty estimates.
  • MC-dropout can be used for data augmentation in uncertainty estimation: Leveraging MC-dropout with VAEs helps distill uncertainty information from more complex models into lightweight conformal bands.
  • A novel joint training loss function is key for efficient CJP: The proposed loss function in CJP effectively balances coverage, sharpness, and computational efficiency by integrating interval scoring and calibration objectives.
  • The CJP method demonstrates remarkable adaptability: It shows resilience to different noise types and can even adapt to reduced model capacity (e.g., switching from ResNet34 to MobileNetV2) while maintaining meaningful uncertainty estimates.

Key Results

The paper presents qualitative comparisons of the four methods in Figure 5, highlighting differences in average interval length, contiguity, precision, accuracy, coverage, and adaptiveness.

  • CQR (Sec. III-A): Produces rectangular uncertainty regions, often overly conservative, with a relatively high average interval length (e.g., 29.12 for x-coordinate).
  • CSP (Sec. III-B): Offers non-continuous uncertainty boundaries and has a much smaller average interval length (e.g., 3.94 for x-coordinate), but struggles with multivariate data.
  • MCQR w/ MC-dropout (Sec. IV-A): Achieves very tight uncertainty estimates (e.g., 0.01 for x-coordinate) but can be overly optimistic and is highly data-dependent, with high computational complexity.
  • CJP (Sec. IV-B): Demonstrates optimal uncertainty estimates with tunability. It achieves a good balance between coverage and sharpness, with moderate average interval lengths (e.g., 8.34 for x-coordinate) that adapt to scene conditions and noise. Figure 6 further shows CJP's resilience to various noise types (Gaussian, Salt&Pepper, Speckle) and its ability to adapt to a less powerful feature extractor (MobileNetV2).

Limitations and Future Work

  • Data Dependency of MCQR: The MCQR method's performance can be inconsistent across different datasets, indicating a strong reliance on the specific data distribution.
  • Computational Complexity of MCQR: This method has the highest computational complexity among the four proposed approaches.
  • Multivariate Data Handling in CSP: CSP faces challenges in effectively handling multivariate data, limiting its ability to accurately estimate true uncertainty in complex scenarios.
  • Future Work: The paper suggests that the CJP method, with its balance of accuracy, adaptability, and computational efficiency, is a promising direction for future research in uncertainty-aware VO for edge robotics. The authors also imply that the framework's robustness across different datasets and noise conditions makes it suitable for a wide range of applications.

Summary updated

BibTeX
@inproceedings{stutts2023lightweight,
  author = {Alex C. Stutts and Danilo Erricolo and Theja Tulabandhula and Amit Ranjan Trivedi},
  title = {{Lightweight, Uncertainty-Aware Conformalized Visual Odometry}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2023},
  doi = {10.1109/iros55552.2023.10341924},
}
Figure