Thank you for your excellent work on CLARE! I am trying to reproduce the reported ~75% success rate on LIBERO-10 but encountering issues.
Setup:
Pretrained model: continuallearning/dit_flow_mt_libero_90_pretrain
Script: bash bash/dit_dec.sh (with policy.path updated)
Benchmark: LIBERO-10
Problem:
After training, both eval/pc_success and eval/avg_sum_reward_Libero remain at 0. I also noticed the previous issue (#XX) about the success detection bug, but in my case the reward is also 0, suggesting something more fundamental might be off.
Questions:
Could you confirm the exact steps and hyperparameters to reproduce the LIBERO-10 results?
Is there an updated evaluation script that fixes the success detection bug?
Are there specific environment versions (MuJocoCo, LIBERO, etc.) required?
Any guidance would be greatly appreciated!
Thank you for your excellent work on CLARE! I am trying to reproduce the reported ~75% success rate on LIBERO-10 but encountering issues.
Setup:
Pretrained model: continuallearning/dit_flow_mt_libero_90_pretrain
Script: bash bash/dit_dec.sh (with policy.path updated)
Benchmark: LIBERO-10
Problem:
After training, both eval/pc_success and eval/avg_sum_reward_Libero remain at 0. I also noticed the previous issue (#XX) about the success detection bug, but in my case the reward is also 0, suggesting something more fundamental might be off.
Questions:
Could you confirm the exact steps and hyperparameters to reproduce the LIBERO-10 results?
Is there an updated evaluation script that fixes the success detection bug?
Are there specific environment versions (MuJocoCo, LIBERO, etc.) required?
Any guidance would be greatly appreciated!