Problems with OVMM result reproduction

[Abc12128]

Hello.
I tried to reproduce the results from the original paper on my server and got the following data.
Averaged metrics
= = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = =
episode_count: 1199.0
does_want_terminate: 0.4061718098415346
num_steps: 874.3436196830693
find_object_phase_success: 0.6688907422852377
pick_object_phase_success: 0.6205170975813178
find_recep_phase_success: 0.5187656380316931
overall_success: 0.14595496246872394
partial_success: 0.48853211009174313
= = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = =
100% | █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ █ | 1199/1199 [12:14:56<00:00, 36.78s/it]
Metrics:
{'episode_count': 1199.0, 'does_want_terminate': 0.4061718098415346, 'num_steps': 874.3436196830693, 'find_object_phase_success': 0.6688907422852377, 'pick_object_phase_success': 0.6205170975813178, 'find_recep_phase_success': 0.5187656380316931, 'overall_success': 0.14595496246872394, 'partial_success': 0.48853211009174313}
But in my results, overall_success is always lower than in the original paper (I chose the baseline configuration in the red box below). I changed --baseline_config_path to rl_agent.yaml and GROUND_TRUTH_SEMANTICS:1) in hssd_eval.yaml.

Do you know the cause of this result?