Alpha-Ordered Gradients

Code for "Do Differentible Simulators Give Better Policy Gradients"?

Setup

Add the git repo to your PYTHONPATH. Then test by import alpha_gradient

Running Examples

We provide multiple examples that can be run.

To visualize the per-coordinate bias and variance on simple one-step examples,

• BallWithWall example: python3 examples/ball_with_wall/alpha_coordinate_sweep.py
• Pivot example: python3 examples/pivot/alpha_coordinate_sweep.py

We include some trajectory optimization examples.

• Friction example: python3 examples/friction/friction_test.py
• Pushing with stiffness 10: python3 examples/curling/curling_10.py
• Pushing with stiffness 1000: python3 examples/curling/curling_1000.py
• Robot motion planning: python3 examples/motion_planning/roomba_test.py

Closed-loop policy optimization examples are:

• Finite-Horizon Static-Policy LQR: python3 examples/linear_system/linear_test.py
• Tennis: python3 examples/breakout/run_bc_policyopt.py