Simple inverse reinforcement learning using RRT as trajectory sampling algorithm. The cost function is linear in features and the gradient at each iteration is the expected difference in agent and expert feature vector. The RRT implementation is from Python Robotics.

cd scripts
python3 irl_rrt_train.py --visualize_trajs

The task is to navigate from start (0, 0) to goal (1, 1). There are two features for each state: 1. distance to goal at (1, 1) and 2. distance to an attractor at (0, 1). The true weights are (0.1, 0.9), which is used by the expert to generate demonstrations. As a result, the sampled expert trajectories bend upwards.

The learned weight is initialized at (0.9, 0.1) which results in almost straight paths from start to goal.

After training the sampled agent paths look like

The figure below shows the learned and true weights during training.

We can also use a neural network to model the cost function instead of the linear feature model used above.

cd scripts
python3 irl_rrt_nn_train.py --visualize_trajs

After a few iterations, we can get sampled agent paths like