Task planning framework supports collaborative furniture assembly between humans and robots

Ensuring that robots can effectively collaborate with humans in real-world environments is essential before deploying them at scale. While some robotic systems already interact with human agents on a daily basis, for example in partially automated industrial and manufacturing facilities, human-robot collaboration on everyday tasks remains rare.

Researchers from the University of Padua and Mitsubishi Electric Research Laboratories (MERL) in Cambridge have developed a framework that allows for the planning of tasks involving human-robot collaboration. The framework, presented in a pre-published article on the site arXiv server, focuses specifically on tasks that involve the collaborative assembly of complex systems with various underlying components, such as furniture.

The researchers named their framework DECAF, which stands for Discrete-Event based Collaborative Human-Robot Assembly Framework for furniture. DECAF has several underlying components, including a discrete-event Markov decision process (DE-MDP) model, an HTM description of the assembly process, and a Bayesian interference module.

“The human is characterized as an uncontrollable agent, which implies, for example, that the agent is not bound by a pre-established sequence of actions and instead acts according to its own preferences,” Giulio Giacomuzzo, Matteo Terreran and their colleagues write in their paper. “Meanwhile, the task planner reactively calculates the optimal actions so that the collaborative robot can efficiently complete the entire assembly task in the shortest possible time.”

Thanks to the new framework developed, the collaborative assembly process extends over several stages. First, the robot observes the actions performed by the human agent, via a camera or other sensors.

Based on these observations, the DECAF model plans actions for the robot that will maximize the efficiency of the robot-human team in performing the assembly tasks at hand, while adapting those actions in response to unpredictable events. The team modeled the assembly of furniture or other objects using a mathematical framework often used to frame specific decision-making processes, known as DE-MDP.

“We formalize the problem as DE-MDP, a comprehensive framework that integrates a variety of asynchronous behaviors, human mindset changes, and failure recovery as stochastic events,” Giacomuzzo, Terreran, and their colleagues wrote.

“Although the problem could theoretically be solved by constructing a graph of all possible actions, such an approach would be limited by computational limitations. The proposed formulation offers an alternative solution using reinforcement learning to derive an optimal policy for the robot.”

Essentially, the DE-MDP model is used to decompose an assembly task and identify the optimal actions that would allow the robot to tackle it efficiently in collaboration with a human agent. The second component of the DECAF framework, namely the HTM model, encodes the interdependence between various subtasks, thereby facilitating the planning process.

Finally, the team integrated a module based on a statistical approach known as Bayesian interference, which is typically used to continuously update the probability that a given hypothesis is true as new information becomes available. This module allows the framework to monitor the actions of the human agent and use them to predict the intentions of a human agent.

The researchers evaluated DECAF in a series of tests, both simulated and real-world. For the real-world experiment, 10 adult participants were asked to assemble a chair purchased from IKEA in collaboration with a 7-degree-of-freedom robotic manipulator (i.e., Franka Emika’s Panda Arm).

The results of the team’s initial tests have proven very promising. In simulations, the DECAF framework outperformed standard planning policies, while in real-world experiments it appeared to improve the efficiency and quality of human-robot collaboration.

“In the future, we plan to include other optimal metrics beyond runtime, such as human safety, action correlation, and human usability,” the researchers wrote.

© 2024 Science X Network