Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Uncertainty-driven Exploration Strategies for Online Grasp Learning

Yitian Shi, Philipp Schillinger, Miroslav Gabriel, Alexander Qualmann, Zohar Feldman, Hanna Ziesche, Ngo Anh Vien

TL;DR

The online learning algorithm with an effective exploration strategy can significantly improve its adaptation performance to unseen environment settings and various uncertainty estimation schemes based on Bayesian uncertainty quantification and distributional ensembles are proposed.

Abstract

Existing grasp prediction approaches are mostly based on offline learning, while, ignoring the exploratory grasp learning during online adaptation to new picking scenarios, i.e., objects that are unseen or out-of-domain (OOD), camera and bin settings, etc. In this paper, we present an uncertainty-based approach for online learning of grasp predictions for robotic bin picking. Specifically, the online learning algorithm with an effective exploration strategy can significantly improve its adaptation performance to unseen environment settings. To this end, we first propose to formulate online grasp learning as an RL problem that will allow us to adapt both grasp reward prediction and grasp poses. We propose various uncertainty estimation schemes based on Bayesian uncertainty quantification and distributional ensembles. We carry out evaluations on real-world bin picking scenes of varying difficulty. The objects in the bin have various challenging physical and perceptual characteristics that can be characterized by semi- or total transparency, and irregular or curved surfaces. The results of our experiments demonstrate a notable improvement of grasp performance in comparison to conventional online learning methods which incorporate only naive exploration strategies. Video: https://youtu.be/fPKOrjC2QrU

Uncertainty-driven Exploration Strategies for Online Grasp Learning

TL;DR

The online learning algorithm with an effective exploration strategy can significantly improve its adaptation performance to unseen environment settings and various uncertainty estimation schemes based on Bayesian uncertainty quantification and distributional ensembles are proposed.

Abstract

Existing grasp prediction approaches are mostly based on offline learning, while, ignoring the exploratory grasp learning during online adaptation to new picking scenarios, i.e., objects that are unseen or out-of-domain (OOD), camera and bin settings, etc. In this paper, we present an uncertainty-based approach for online learning of grasp predictions for robotic bin picking. Specifically, the online learning algorithm with an effective exploration strategy can significantly improve its adaptation performance to unseen environment settings. To this end, we first propose to formulate online grasp learning as an RL problem that will allow us to adapt both grasp reward prediction and grasp poses. We propose various uncertainty estimation schemes based on Bayesian uncertainty quantification and distributional ensembles. We carry out evaluations on real-world bin picking scenes of varying difficulty. The objects in the bin have various challenging physical and perceptual characteristics that can be characterized by semi- or total transparency, and irregular or curved surfaces. The results of our experiments demonstrate a notable improvement of grasp performance in comparison to conventional online learning methods which incorporate only naive exploration strategies. Video: https://youtu.be/fPKOrjC2QrU
Paper Structure (26 sections, 5 equations, 6 figures)

This paper contains 26 sections, 5 equations, 6 figures.

Table of Contents

  1. Introduction
  2. Related works
  3. Modern Robot Grasping Methods
  4. Deep RL-based Grasping Methods
  5. Uncertainty Estimation in Deep RL
  6. Uncertainty-driven Offline-to-Online Robotic Grasp Learning
  7. Problem Formulation
  8. Ensemble of Probabilistic ConvSAC
  9. Gaussian-based Uncertainty Estimation
  10. Quantile-based Distributional Learning
  11. Data-efficient Online Training
  12. Multi-processed Online Learning
  13. UCB-based Exploration Policy
  14. Online Training with Data Buffer
  15. Experiments
  16. ...and 11 more sections

Figures (6)

  • Figure 1: Given an image of a picking scene (1st column), every 500 training steps our MV-ConvSACs predicts a grasp reward map (2nd column) and a normalized uncertainty map (3rd/4th columns).
  • Figure 2: Ensemble of probabilistic ConvSAC network architecture: MV-ConvSACs and QR-ConvSACs.
  • Figure 3: Grasping experiment setup
  • Figure 4: Ablation on uncertainty exploration strategies for MV-ConvSACs and QR-ConvSACs
  • Figure 5: Online training steps for MV- and QR-ConvSACs
  • ...and 1 more figures