add policy client for GR00T N1.6

[12e21]

Summary:
add Gr00t16ServicePolicyClient to support GR00t N1.6 new policy server api #103 , compatibility with other policy types
, use as:

python scripts/evaluation/policy_inference.py \
    --task=LeIsaac-SO101-PickOrange-v0 \
    --eval_rounds=10 \
    --policy_type=gr00tn1.6 \
    --policy_host=localhost \
    --policy_port=5555 \
    --policy_timeout_ms=5000 \
    --policy_action_horizon=16 \
    --policy_language_instruction="Pick up the orange and place it on the plate" \
    --device=cuda \
    --enable_cameras

Verification:
Tested locally: Confirmed successful deploy of finetuned GR00T N1.6 model

[EverNorif]

@12e21 Thank you for your contribution.
I don’t have the resources to test the GR00T v1.6 related changes this week, will carry out the testing after one week. :)

[EverNorif]

@12e21 Hi, I’m currently testing this PR, but I ran into an issue when fine-tuning gr00t n1.6. I’m fine-tuning on a 4090, but even when I set global_batch_size = 1, I still get a CUDA Out of Memory error. Could you share your fine-tuning configuration?

[12e21]

@EverNorif I tried finetuning with 4090 24GB, indeed OOM with batch size 1; on A100 or A6000, it works with about 40GB VRAM

[12e21]

@EverNorif I can share my config and upload finetuned policy model to huggingface, so you can test it

[EverNorif]

@EverNorif I can share my config and upload finetuned policy model to huggingface, so you can test it

Okay, that would be better.

[12e21]

@EverNorif I have upload configs, scripts, and checkpoint, find at 12e21/gr00t_n1d6_leisaac_pick_orange