Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ELLIPSE: Evidential Learning for Robust Waypoints and Uncertainties

Zihao Dong, Chanyoung Chung, Dong-Ki Kim, Mukhtar Maulimov, Xiangyun Meng, Harmish Khambhaita, Ali-akbar Agha-mohammadi, Amirreza Shaban

TL;DR

A lightweight domain augmentation procedure that synthesizes plausible viewpoint/pose variations without collecting additional demonstrations to reduce covariate-shift-induced overconfidence under viewpoint and pose perturbations near expert trajectories and improve uncertainty reliability under environment/domain shift.

Abstract

Robust waypoint prediction is crucial for mobile robots operating in open-world, safety-critical settings. While Imitation Learning (IL) methods have demonstrated great success in practice, they are susceptible to distribution shifts: the policy can become dangerously overconfident in unfamiliar states. In this paper, we present \textit{ELLIPSE}, a method building on multivariate deep evidential regression to output waypoints and multivariate Student-t predictive distributions in a single forward pass. To reduce covariate-shift-induced overconfidence under viewpoint and pose perturbations near expert trajectories, we introduce a lightweight domain augmentation procedure that synthesizes plausible viewpoint/pose variations without collecting additional demonstrations. To improve uncertainty reliability under environment/domain shift (e.g., unseen staircases), we apply a post-hoc isotonic recalibration on probability integral transform (PIT) values so that prediction sets remain plausible during deployment. We ground the discussion and experiments in staircase waypoint prediction, where obtaining robust waypoint and uncertainty is pivotal. Extensive real world evaluations show that \textit{ELLIPSE} improves both task success rate and uncertainty coverage compared to baselines.

ELLIPSE: Evidential Learning for Robust Waypoints and Uncertainties

TL;DR

A lightweight domain augmentation procedure that synthesizes plausible viewpoint/pose variations without collecting additional demonstrations to reduce covariate-shift-induced overconfidence under viewpoint and pose perturbations near expert trajectories and improve uncertainty reliability under environment/domain shift.

Abstract

Robust waypoint prediction is crucial for mobile robots operating in open-world, safety-critical settings. While Imitation Learning (IL) methods have demonstrated great success in practice, they are susceptible to distribution shifts: the policy can become dangerously overconfident in unfamiliar states. In this paper, we present \textit{ELLIPSE}, a method building on multivariate deep evidential regression to output waypoints and multivariate Student-t predictive distributions in a single forward pass. To reduce covariate-shift-induced overconfidence under viewpoint and pose perturbations near expert trajectories, we introduce a lightweight domain augmentation procedure that synthesizes plausible viewpoint/pose variations without collecting additional demonstrations. To improve uncertainty reliability under environment/domain shift (e.g., unseen staircases), we apply a post-hoc isotonic recalibration on probability integral transform (PIT) values so that prediction sets remain plausible during deployment. We ground the discussion and experiments in staircase waypoint prediction, where obtaining robust waypoint and uncertainty is pivotal. Extensive real world evaluations show that \textit{ELLIPSE} improves both task success rate and uncertainty coverage compared to baselines.
Paper Structure (18 sections, 11 equations, 5 figures, 2 tables)

This paper contains 18 sections, 11 equations, 5 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Uncertainty Quantification
  4. Imitation Learning and Covariate Shift
  5. Synthesizing Demonstrations
  6. Interactive Imitation Learning
  7. Placement of This Work
  8. Method
  9. Learning Waypoints with Deep Evidential Regression
  10. Domain Augmentation via Synthesizing Novel Viewpoints
  11. Calibrating Uncertainty with Isotonic Regression
  12. Integrating Learned Uncertainty with Motion Planner
  13. Experiments
  14. Experiment Setup
  15. Stair Climbing Success Rate on Unseen Stairs
  16. ...and 3 more sections

Figures (5)

  • Figure 1: Deployment environments (\ref{['sec:exp:setup']}). Such environments are especially challenging due to limited sensor FOV, narrow passageways (landings), and invisible stair boundaries (hollow handrails and glass). The ability to navigate stairs is pivotal for robots to efficiently explore multi-floor structures.
  • Figure 2: Overall pipeline of ELLIPSE. The network is trained offline with domain augmentation to improve robustness of the waypoints and uncertainties (\ref{['sec:method:aug']}). During inference, the predicted waypoint distributions (\ref{['sec:method:der']}) are recalibrated using scales obtained from Isotonic Regression (\ref{['sec:method:isotonic']}). A Mahalanobis-distance-based uncertainty-aware MPPI planner runs on a separate thread and tracks a pool of waypoints (at high frequency) while relaxing constraints on uncertain waypoints (\ref{['sec:method:mppi']}).
  • Figure 3: Uncertainty is calibrated on train yet overconfident on test.
  • Figure 4: Timelapse and trajectories of the Spot traversing CLF using different variants of ELLIPSE. (a)(b): Without the domain augmentation, both variants crash into handrails due to compounding error. (c)(d): With the domain augmentation, both variants completes the run without help, and stay closer to stair center.
  • Figure 5: Qualitative comparison of MPPI planning variants, overlaid with unit-level uncertainty ellipses (accepted, relaxed; current-step predictions are highlighted with black edges). Mahalanobis+Hist remains close to confident predictions, whereas the other variants can be dominated by highly uncertain waypoints, leading to potentially unsafe behavior under disturbances.