Learning-Based Multi-Robot Lane Navigation: Scalable Trajectory Prediction using Neural Networks

Problem Statement & Motivation

Accurate trajectory prediction is crucial for safe and efficient robot navigation. While high-fidelity simulators like Webots offer reliable results, they lack scalability for real-time deployment or multi-agent scenarios. The goal of this project is to implement a scalable neural network model that can generate robot trajectories efficiently, with acceptable precision.

The task involves navigating a cyclic path with a variable number of lanes, introducing additional planning complexity.

Our Method

We investigated a range of machine learning techniques to model robot motion:

1. Environment Setup:
   • The navigation task was simulated in Webots.
   • The environment includes a cyclic lane path with adjustable width and complexity.
2. Trajectory Prediction:
   • Input: initial robot position.
   • Output: incremental prediction of next positions using the neural network.
   • The model is autoregressive: each predicted step becomes the input for the next.
3. Learning Approaches:
   • MLP, RNN, and GNN: struggled to maintain trajectory accuracy.
   • Reinforcement Learning (PPO): yielded stable behaviors.
   • Imitation Learning: learned from expert trajectories.
   • Combined RL + Imitation Learning: achieved best performance in terms of accuracy and scalability.
4. Simulation Loop:
   • The predicted movement is integrated step-by-step (xₜ → xₜ₊₁).
   • The learned policy generalizes over multiple steps.

Evaluation & Results

• The best-performing model used imitation learning guided by reinforcement learning, balancing precision with efficiency.

Observations

• Training time is substantial and environment-specific.
• The learned controller is scalable and can potentially extend to multi-agent settings with further training.