Wednesday, September 18, 2024

DECAF human-robot furniture assembly

The Framework Supports Human-Robot Collaboration by Streamlining Task Planning for Furniture Assembly

Importance of Effective Human-Robot Interaction

Furniture assembly

Ensuring effective Human-Robot interaction in real-world environments is essential for broad deployment. Although some robotic systems collaborate with humans in partially automated settings, everyday tasks still see limited collaboration.

Introduction of DECAF

Development and Purpose

A team of researchers from the University of Padova and Mitsubishi Electric Research Laboratories (MERL) in Cambridge has created a Task-Planning framework for Human-Robot collaboration. Detailed in a preprint on arXiv, the framework is designed for complex assembly tasks involving multiple components, such as furniture.

Framework Components

The researchers introduced their framework as DECAF---Discrete-Event based Collaborative Human-Robot Assembly Framework for furniture. It features several core components, such as a discrete-event Markov decision process (DE-MDP), an HTM assembly description, and a Bayesian inference component.

Functionality and Process

Human-Agent Dynamics

"The human is defined as an uncontrollable agent, acting according to personal preferences rather than a predetermined sequence, as noted by Giulio Giacomuzzo, Matteo Terreran, and their co-authors. Concurrently, the task planner computes the robot's optimal actions to efficiently complete the assembly task in minimal time."

Collaborative Assembly Process

In the newly established framework, the collaborative assembly process is divided into several stages. The first step involves the robot observing the human agent's actions using a camera or similar sensors.

Planning and Adjustment

The DECAF model utilizes these observations to plan optimal actions for the robot, aimed at maximizing the efficiency of the Robot-Human team in assembly tasks, and adjusting for unforeseen circumstances. The team modeled the assembly process for furniture and other objects using the DE-MDP framework, a common tool in decision-making contexts.

Technical Details

DE-MDP Framework

According to Giacomuzzo, Terreran, and their colleagues, the problem is formalized within the DE-MDP framwrok, which encompasses various asynchronous behaviros, human decision changes, and failure recovery as stochastic events.

Computational Approach

"While it is theoretically possible to tackle the problem by creating a graph of all potential actions, this method would be limited by computational constraints. The proposed approach offers a viable alternative by employing Reinforcement Learning to determine an optimal policy for the robot."

Task Dissection and Planning

The DE-MDP model fundamentally dissects an assembly task to determine the optimal actions for the robot to perform efficiently in conjunction with a human agent. Complementing this, the HTM model within the DECAF framework encodes the dependencies among sub-tasks, thereby streamlining the planning process.

Bayesian Inference Module

Finally, a module employing Bayesian inference was integrated into the system. This statistical approach, which updates the probability of a hypothesis as more data becomes available, allows the framework to monitor human actions and infer their intentions.

Evaluation and Future Directions

Testing and Results

The researchers assessed the DECAF framework through a series of tests, including both simulations and real-world applications. In the real-world trial, ten adult participants were tasked with assembling an IKEA chair in collaboration with a 7-DoF robotic arm, specifically the Franka Emika Panda.

Initial Findings

The initial test results were highly encouraging. In simulations, the DECAF framework demonstrated superior performance compared to conventional planning policies, and in real-world experiments, it enhanced both the efficiency and quality of human-robot collaboration.

Future Enhancements

Future enhancements will involve incorporating additional metrics beyond execution time, such as human safety, action correlation, and ergonomics, as outlined by the researchers.

Source

Labels: , , ,

Friday, July 12, 2024

human-robot interaction risk analysis

What level of risk aversion is observed among humans when interacting with robots?
What are the preferred methods for human-robot interaction in crowded environments?
Which algorithms should roboticists employ to program robots for effective human interaction?
These questions were the focus of a study conducted by mechanical engineers and computer scientists at the University of California San Diego, recently presented at the International Conference on Robotics and Automation (ICRA) 2024 in Japan.
"This study represents the first known investigation into robots that infer human risk perception for intelligent decision-making in everyday contexts," stated Aamodh Suresh, the study's first author, who earned his Ph.D. under Professor Sonia Martinez Diaz in UC San Diego's Department of Mechanical and Aerospace Engineering. He now serves as a postdoctoral researcher at the U.S. Army Research Lab.
"Our goal was to develop a framework to understand human risk aversion in interactions with robots," said Angelique Taylor, the study's second author, who earned her Ph.D. in Computer Science and Engineering under Professor Laurel Riek at UC San Diego. Taylor is now a faculty member at Cornell Tech in New York.
The team utilized behavioral economics models, deliberating over their selection amidst the pandemic. Consequently, researchers implemented an online experiment to ascertain the optimal model for their study.
The study involved STEM undergraduate and graduate students who participated in a simulation as Instacart shoppers. They were presented with three route options to reach the milk aisle in a grocery store, each spanning between five to 20 minutes. Certain routes exposed them to individuals infected with COVID-19, including one path featuring a severe case.
Each path also carried distinct risks of exposure to individuals with COVID-19 coughing incidents. The shortest route notably exposed participants to a higher concentration of severely ill individuals. Nevertheless, shoppers were motivated by rewards tied to expedient goal attainment.
The researchers were astonished to find a consistent underestimation in survey responses regarding individuals' readiness to risk proximity with COVID-19-infected shoppers. "When there's a reward involved, people seem more willing to take risks," remarked Suresh.
Consequently, in programming robots for human interaction, researchers opted to employ prospect theory, a behavioral economics model pioneered by Daniel Kahneman, Nobel laureate in economics (2002). This theory posits that individuals assess potential losses and gains relative to a reference point.

In this context, individuals tend to weigh losses more heavily than gains. For instance, participants in the study chose to receive $450 rather than taking a 50% chance to win $1100. Thus, subjects focused on obtaining a certain reward promptly rather than evaluating the potential risk of contracting COVID-19.

Researchers also queried individuals on their preferences regarding how robots should convey their intentions. Responses encompassed speech, gestures, and touch screens.

Labels: