[Paper Review] Grasp Learning by Sampling from Demonstration

P Zech, Justus Piater|arXiv (Cornell University)|Nov 19, 2016

Robot Manipulation and Learning4 references3 citations

TL;DR

This paper proposes a model-free, active grasp learning method that learns object grasp affordances using only a few user-demonstrated 6D gripper poses and object pose, without requiring geometric shape information. By combining MCMC Kameleon and Generalized Darting Monte Carlo (GDMC), the method samples from a grasp affordance density to discover diverse, high-quality grasps, achieving robust performance even in cluttered environments with minimal prior knowledge.

ABSTRACT

Robotic grasping traditionally relies on object features or shape information for learning new or applying already learned grasps. We argue however that such a strong reliance on object geometric information renders grasping and grasp learning a difficult task in the event of cluttered environments with high uncertainty where reasonable object models are not available. This being so, in this paper we thus investigate the application of model-free stochastic optimization for grasp learning. For this, our proposed learning method requires just a handful of user-demonstrated grasps and an initial prior by a rough sketch of an object's grasp affordance density, yet no object geometric knowledge except for its pose. Our experiments show promising applicability of our proposed learning method.

Motivation & Objective

• To address the challenge of robotic grasp learning in cluttered, uncertain environments where accurate object geometry is unavailable.
• To reduce reliance on object shape or geometric modeling in grasp learning, enabling robust performance in real-world scenarios.
• To develop a method that learns grasp affordance densities using only a few demonstrated grasps and object pose.
• To enable efficient exploration of the 6D gripper pose space through stochastic optimization, avoiding shape-based or analytic computation.

Proposed method

• The method uses Metropolis-Hastings Markov Chain Monte Carlo (MCMC) sampling to explore the 6D gripper pose space relative to a canonical object pose.
• MCMC Kameleon is employed to adaptively learn the covariance structure of the grasp affordance density, enabling sampling from complex, non-linear distributions.
• Generalized Darting Monte Carlo (GDMC) enhances mixing between modes by guiding proposals toward existing grasp modes, improving convergence and exploration.
• A rough sketch of the grasp affordance density is used as an initial prior, while demonstrated grasps provide initial modes for the sampling process.
• The sampling process is biased via a parameter $ c $, which controls the gradient step size in translational dimensions, balancing exploration and exploitation.
• The method operates under the assumption that object pose is known, avoiding the need for object pose estimation.

Experimental results

Research questions

• RQ1 Can grasp affordance densities be effectively learned without geometric object models or shape information?
• RQ2 How well can a few demonstrated grasps guide the discovery of diverse, high-quality grasps in a 6D pose space?
• RQ3 What is the impact of different biasing strategies (impartial, weak, strong) on the sampling efficiency and grasp diversity?
• RQ4 How does the parameter $ c $, controlling translational exploration, affect the performance and robustness of the grasp learning process?
• RQ5 Can MCMC-based sampling with adaptive covariance and mode-aware proposals outperform traditional shape- or trajectory-based grasp learning in cluttered settings?

Key findings

• The proposed method consistently discovered hundreds of grasps across all tested objects (pitcher, pan, plate), even with minimal geometric prior knowledge.
• With a weak or strong bias, the method achieved mean grasp counts of 216 ± 46 (pan), 180 ± 46 (plate), and 143 ± 48 (pitcher), demonstrating robustness to biasing strategies.
• The optimal value of the control parameter $ c $ for exploration was found to lie between 0.1 and 0.25, with slight extension to 0.15 for elongated objects like the pan.
• Overly high $ c $ values led to poor performance due to excessive translational moves that caused the sampler to overshoot the object.
• Dispersion plots confirmed that the method effectively explored diverse regions of the grasp space, with convex hull areas increasing with appropriate $ c $, indicating strong exploration behavior.
• The method demonstrated that grasp learning can be performed effectively without object geometry, relying only on pose and a few demonstrations, proving the feasibility of 'blind' grasp learning.