Real Evaluations Tractability using Continuous Goal-Directed Actions in Smart City Applications.

Area of Science:

• Robotics and Human-Robot Interaction
• Artificial Intelligence and Machine Learning
• Smart City Technologies

Background:

• Smart City applications require intuitive robot interaction with non-expert users.
• Traditional robot imitation frameworks use geometric trajectories, limiting representation of visual features.
• Continuous Goal-Directed Actions (CGDA) offer a feature-agnostic approach but require extensive computation via Evolutionary Algorithms (EA).

Purpose of the Study:

• To investigate the feasibility of performing EA evaluations directly in real-world robot scenarios for CGDA.
• To reduce the number of evaluations needed for CGDA in dynamic, complex environments.
• To compare the effectiveness of different optimization strategies within the CGDA framework.

Main Methods:

• Implemented and compared Particle Swarm Optimization (PSO) variants (naïve PSO, FI-PSO, AFFG-PSO) within the CGDA framework.
• Introduced geometrical and velocity constraints into the CGDA framework as a second optimization approach.
• Evaluated both approaches on 'wax' and 'paint' action tasks, common CGDA use cases.

Main Results:

• Both proposed approaches demonstrated a significant reduction in the number of required EA evaluations.
• PSO-based methods and constraint incorporation proved effective in optimizing CGDA.
• Direct real-world evaluations are tractable and offer advantages over simulation-based methods.

Conclusions:

• The study successfully demonstrates methods to reduce computational load for CGDA in real-world robotics.
• These advancements facilitate more efficient robot programming for smart city applications.
• Future work can explore further optimization techniques and broader application of CGDA.