Results expectation about ecrl HOT 1 CLOSED

Hi Hanna,

1. Training through main.py: The Eval Episode Video and Goal Image are recorded every config['Training']['epochEpisodes'] when evaluating the policy during training. The other videos and goals are recorded every config['WANDB']['episodeVisFreq'] . The current config, if you haven't modified it, is 800 episodes for both. Notice that the Eval video displays a policy that seems less stochastic, this is because there is no exploration noise added when evaluating.

Evaluation with policy_eval.py: This script evaluates a chosen policy and prints the performance statistics as well as saves a single rollout video + goal in order to visualize the policy behavior.

2. The code will run config['Training']['totalTimesteps'] timesteps in the environment.

3. The inputs to the policy and Q-function neural networks are normalized using running statistics (running mean and std). The obs_rms_std and obs_rms_mean graphs record a mean of these across the object feature dimension and it is mostly meant to identify anomalies such as a sudden very large change in statistics that might indicate a bug. What you should expect in these graphs is relatively quick convergence to some value, 'obs_rms_std' shouldn't be too close to zero. The graphs that indicate performance are eval_goal_achievement_%, mean_success_frac, mean_avg_obj_dist, mean_max_obj_dist and eval_mean_reward. I would also look at 1C_interaction_rate which can signal when the agent starts to learn that it needs to interact with objects in order to achieve its goals.

Let me know if you have any other questions. You can also email me if you prefer, I will be happy to help.

Best,
Dan

from ecrl.