Apex 0.3.0

This release creates a distinction between reference trajectory and no-reference trajectory environments (CassieTrajEnv and CassieEnv respectively), but leaves other attributes such as learning PD gains, dynamics randomization, and using a full or minimal input as arguments to the environment's constructor.

Instead of clock_based and phase_based, environments now have a command_profile attribute which specifies the type of command input to the policy. This can be clock or phase, or even traj in the case of CassieTrajEnv. Another new attribute, input_profile, specifies the size and composition of the policy's input. This can be full or min. Naturally the number of choices for command_profile and input_profile can be readily expanded as research progresses.

Other notable features:

• Policy comparison Test
• Playground environment for running autonomous missions
• Custom terrain via PNG height maps
• Comprehensive 5K test
• Live plot when evaluating phase_based policies.
• More insightful PPO logging info (optimize and sample times)

Apex 0.2.0

Features:

• highly customizable unified cassie environment
• reference trajectories (Agility controller, ASLIP model)
• several distributed RL algorithms with Ray
• mirror symmetry loss for PPO
• common entry-point for training and testing policies
• tensorboard logging
• tools for evaluating trained policies