rskill-rldx1-pt-nf4

Pretrain (foundation) checkpoint for the RLDX-1 collection. Listed in-tree so contributors who want to finetune RLDX-1 on a new embodiment can pin the upstream revision.

This is a sibling of rldx1-ft-libero-nf4; that README owns the canonical architecture, license, auto-managed sidecar lifecycle, and NF4 quantization documentation for every member of the RLDX-1 family. Read it first.

This is not a deployable s1 policy on its own — RLDX-1-PT has no per-suite finetune and is not benchmarked. Use it via:

# Finetune from the PT checkpoint for your embodiment (upstream workflow):
python -m rldx.training.finetune \
    --model-path RLWRLD/RLDX-1-PT \
    --dataset <your-LeRobotDataset> \
    --embodiment-tag <your-embodiment>

# Then re-package the resulting checkpoint as a new rskill mirroring
# the rldx1-ft-libero-nf4 layout.

Upstream: https://huggingface.co/RLWRLD/RLDX-1-PT