Neural network approaches to dynamic collision-free trajectory generation.

Area of Science:

• Robotics
• Computational Neuroscience
• Artificial Intelligence

Background:

• Generating dynamic, collision-free trajectories for robots in non-stationary environments is a significant challenge.
• Existing methods often require extensive computation, prior environmental knowledge, or learning procedures.

Purpose of the Study:

• To develop a computationally efficient, biologically inspired neural network for real-time, dynamic, collision-free trajectory generation.
• To demonstrate the model's effectiveness in various complex scenarios without explicit optimization or learning.

Main Methods:

• Utilizing a topologically organized neural network with neurons governed by shunting or additive equations.
• Mapping the robot's state space (Cartesian or joint) to the neural network's dynamic activity landscape.
• Employing local lateral connections for neuron interaction.

Main Results:

• The neural network generates optimal, collision-free trajectories in real-time by leveraging its dynamic activity landscape.
• The approach does not require explicit free-space searching, global cost function optimization, prior environmental knowledge, or learning.
• The system's stability is assured by a Lyapunov function candidate, and it exhibits robustness to parameter variations.

Conclusions:

• The proposed biologically inspired neural network offers an efficient and effective solution for dynamic collision-free trajectory generation.
• The model's ability to handle non-stationary environments and complex obstacle avoidance demonstrates its practical applicability in robotics.