[BUG] preprocess code produce the different shape retargetting results

[Kashu7100]

The resulting files from retargetting commands (below) have different dimensions from the provided files.
Specifically, those generated has temporal length of 200 but the one provided in assets has 600+.
When I run the RL with those generated, the First dim size mismatch occurs in load_demo in RewardModule (because it loads the 200 length demo and slice it from 30).
With modification on the frame_start and frame_end in the load_contact_retarget_data (frame_start=frame_start, frame_end=frame_end -> frame_start=0, frame_end=len(demo_data["obj_pos"])), I could load the generated files but the RL results are much worse (and eval trials were unsuccessful unlike the policy trained with provided one).

• retargeter_results/${HAND}/${OBJ}_use_01_vector.npy
• retargeted/${HAND}/s01/${OBJ}_use_01_vector_para.pt
• contact_retarget/${HAND}/s01/${OBJ}_use_01.npy

OBJ=box
HAND=inspire
CLIP=box-30-230

python dexmachina/retargeting/parallel_retarget.py --clip $CLIP --hand ${HAND} --control_steps 2000 --save_name para --save -ow

FNAME=dexmachina/assets/arctic/processed/s01/${OBJ}_use_01.npy
python dexmachina/retargeting/map_contacts.py --hand $HAND --load_fname $FNAME 

Could you double check on those preprocess codes as well as arguments to the script so that I can reproduce the retargeting results (.npy and _para.pt) provided?

[MandiZhao]

Hi @Kashu7100 -- thanks for bringing this up! I'll take a closer look at it today. But for now it sounds like the parallel_retarget.py has a slicing step when it's supposed to retarget the entire trajectory -- could you try running with CLIP=box-0-600 and do python dexmachina/retargeting/parallel_retarget.py --clip $CLIP --hand ${HAND} --control_steps 2000 --save_name para --save -ow again?

[nonlinearjunkie]

Hi @Kashu7100 , are the results in orange color from training using the retargeted demo provided by the dexmachina repo?

I tried training the inspire hand:

1. using retargeted demo provided in the assets dir ( represented by brown curve) and
2. run the retargeting script to generate new retargeted data , without any change to hand configurations (represented by orange curve)

Although the new retarget yields better results on curve, still the evaluation seems suboptimal.

Here is the eval video for policy trained using new retargeted data without any change to hand configurations:

inspire_hand-rl_policy-eval.mp4

Do you have any feedbacks on this can be improved and how was your eval demo on inspire hand for use box task?