Online Tool Selection with Learned Grasp Prediction Models

Khashayar Rohanimanesh, Jake Metzger, William Richards, Aviv Tamar

Abstract

In industrial bin-picking, robots must often handle a diverse array of objects requiring different end-effectors. While tool changers provide the necessary flexibility, the time cost associated with swapping tools can significantly degrade throughput. This paper addresses the Grasp Tool Selection Problem (GTSP) by modeling it as a Markov Decision Process (MDP). We propose a real-time planning framework utilizing Model Predictive Control (MPC) and a novel Sparse Tree Search (STS) algorithm. By incorporating "void zones" to account for environment stochasticity and object occlusion, our approach achieves a 50% increase in throughput over standard greedy strategies in real-world trials.

Introduction

Modern industrial automation requires robots to pick and place a chaotic mix of items, ranging from large, flat boxes to small, irregularly shaped consumer goods. To maintain high success rates, robots are often equipped with automatic tool changers that allow them to switch between various vacuum suctions or mechanical grippers. However, tool changing is a double-edged sword: selecting the optimal tool for a specific object increases the probability of a successful grasp, but the act of swapping tools incurs a significant cycle-time penalty.

The challenge lies in online planning. A robot must decide not only which object to pick next but also when to trigger a tool change, despite the fact that many objects remain occluded beneath others. A simple "greedy" strategy—picking the most accessible object with the current tool—often leads to sub-optimal sequences and excessive, time-consuming swaps.

Methodology

Problem Formulation

We formalize the selection process as the Grasp Tool Selection Problem (GTSP), treated as a Markov Decision Process. The goal is to maximize a reward function that balances successful picks against the time cost of tool transitions.

Model Predictive Control and Void Zones

To make the problem tractable in real-time, we employ Model Predictive Control (MPC). Rather than planning for the entire bin, the robot looks ahead over a receding horizon ($H$). To handle the uncertainty of how picking one object affects its neighbors, we introduce Void Zones. When a future pick is planned within the look-ahead window, a geometric radius is established around that pick. The planner assumes that the act of picking will disturb or remove nearby objects, effectively ignoring other grasp proposals in that radius for the remainder of the planning horizon.

Sparse Tree Search (STS)

While the GTSP can be solved optimally using Integer Linear Programming (ILP), ILP is computationally prohibitive for high-speed production environments. We propose Sparse Tree Search (STS), which prunes the decision tree by only considering the top k grasp proposals for each available tool. This heuristic provides near-optimal results while operating orders of magnitude faster than ILP, enabling millisecond-level decision-making.

The $\beta$-TC-Score

Standard throughput (picks per hour) can be a misleading metric, as it may reward aggressive behavior that results in high failure rates or equipment wear. We introduce the $\beta$-TC-Score, a harmonic mean that balances the Pick Success Rate (PSR) and the Tool Consistency Rate (TCR). This metric ensures that the system is evaluated on both its efficiency and its ability to minimize unnecessary tool wear.

Experimental Evaluation

Datasets and Training

The underlying grasp success predictors were trained using a ResNet-101 backbone on a proprietary dataset of approximately 5,000 grasp attempts per tool. For algorithmic validation, we utilized a synthetic simulator that generates grasp maps using Gaussian kernels to mimic object distributions, allowing for rigorous testing of the STS solver against ILP baselines.

Real-World Performance

Testing was conducted on a Fanuc LRMate 200iD/7L industrial arm equipped with 30mm and 50mm vacuum tools. The robot was tasked with clearing a bin filled with a variety of mixed consumer goods.

Key results include:

  • Throughput: The MPC-STS approach achieved 429.75 picks per hour, compared to 317 picks/hr for a naive greedy baseline and 186 picks/hr for a randomized approach.
  • Efficiency: MPC-STS significantly reduced the frequency of tool changes, achieving the highest $\beta$-TC-score (0.6885) among all tested methods.
  • Horizon Optimization: We found that a look-ahead horizon of $H=2$ or $H=3$ is the "sweet spot." Planning further ahead often decreased performance as the deterministic void zone approximation became less accurate over multiple object displacements.

Discussion and Future Work

The primary advantage of this method is the decoupling of learning and planning. By adding a planning layer on top of existing, pre-trained grasp models, the system can be deployed in existing industrial workflows without retraining end-to-end models.

However, the current "void zone" heuristic is a simplified model of bin physics. Future work will explore the integration of deep visual predictive models to simulate bin dynamics more accurately. Additionally, we aim to extend the framework to handle more complex tool sets, including hybrid gripper-vacuum systems and multi-suction cup arrays.

Summary updated

BibTeX
@inproceedings{rohanimanesh2023online,
  author = {Khashayar Rohanimanesh and Jake Metzger and William Richards and Aviv Tamar},
  title = {{Online Tool Selection with Learned Grasp Prediction Models}},
  booktitle = {IEEE International Conference on Robotics and Automation},
  year = {2023},
  doi = {10.1109/icra48891.2023.10160952},
}

Loading related content...

Figure