Performance of cheetah-run task

[jiaweihhuang]

Hi,

I tried to run deployment-efficient experiments to reproduce the results reported in the paper with the following command:

python recursive.py --env cheetah_run --exp_name recursive_example --sub_exp_name BREMEN_demo --param_path configs/params_cheetah_run_offline.json --bc_init --random_seeds 0 --target_kl 0.1 --max_path_length 250 --gaussian 0.1 --const_sampling

After the training is finished, I observed the following evaluation results:

---------------------------
| Iteration    | 399      |
| TotalSamples | 850000   |
| episode_max  | 745      |
| episode_mean | 741      |
| episode_min  | 735      |
---------------------------

However, according to Fig 2 in the original paper (https://arxiv.org/pdf/2006.03647.pdf), the performance of HalfCheetah should be around 6000, which is quite different from the evaluation results.

I wonder the parameter setting specified in the command above is the same as the setting of experiments in this paper? If not, could you let me know which hyper-parameter should be modified in order to reproduce the results reported in the paper? Or maybe there are some other reasons for this performance gap?

Thanks a lot!

[frt03]

@Leonardo-H

cheetah_run uses a different reward function (based on DM Control) from HalfCheetah (based on Open AI Gym MuJoCo). If you reproduce the main results in the paper, you should use HalfCheetah environments (also cheetah_run result is described in Appendix).

[jiaweihhuang]

Got it. Thanks for your reply!