sfujim/td3_bc

Author's PyTorch implementation of TD3+BC, a simple variant of TD3 for offline RL

td3_bc's Issues

use of expl_noise

Couldn't find the usage of expl_noise in actual td3 implementation

Questions about performance metric d4rl_score.

I run the code in halfcheetah-expert-v0, and it seems to work well, but its performance metric d4rl_score is only about 1.1-1.2, and the result of the paper is about 110-120, I am confused. (my mujoco version is 200)

Unable to reproduce results of Antmaze in Table 8

Hi! I have the same question as a previously closed issue. I wasn't able to reproduce results for Antmaze tasks in Table 8. I made the following adjustments in run_experiments.sh, 1. change envs to Antmaze; 2. make normalize False. The .sh file looks like this:

envs=("antmaze-umaze-v0"
"antmaze-umaze-diverse-v0"......)
for ((i=0;i<5;i+=1))
do
for env in ${envs[*]}
do
python main.py
--env $env
--normalize False
--seed $i
done
done

But the normalized scores for last final 10 evaluations and 5 seeds are much lower than the numbers provided in Table 8 inl paper. What's going wrong here? Is there other places I should modify? It would be great to provide how to reproduce Antmaze results. Thanks!

The results in Antmaze

Hi,

I would like to ask the setting about the experiments in Antmaze. Should I need to tune the hyparameters for mujoco locomotion?

I find I cannot reproduce the results about Antmaze in the paper.

Best