rskill-rldx1-mt-droid-nf4

RLDX-1 mid-trained on the DROID Franka real-robot dataset — the only RLDX-1 checkpoint in the collection trained on real-world data — packaged for OpenRAL.

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.

The skill-specific differences from the canonical sibling:

• Upstream weights: hf://RLWRLD/RLDX-1-MT-DROID.
• Training data: real-world Franka teleoperation from the DROID dataset (closest on-hardware-ready checkpoint in the RLDX-1 family — still requires either franka_panda lab access or a closed-loop digital twin to deploy).
• Same NF4-on-backbone, bf16-on-head quantization recipe; same sidecar boot path; same non-commercial license posture.

Hardware deployment: requires both OPENRAL_ACCEPT_NONCOMMERCIAL=1 (research) and a separate vendor agreement with RLWRLD (commercial). The wrapping openral loader refuses commercial activation by default.

Upstream: https://huggingface.co/RLWRLD/RLDX-1-MT-DROID