DecentVLA/droid8_pi05_cent

droid8 π0.5 — cent

Official openpi JAX π0.5, full fine-tune from pi05_base, joint-position DROID recipe: centralised on the union of the 8 clients' data; 100k steps x batch 256 (openpi pi05_full_droid_finetune budget). Same inputs / outputs as openpi's pi05_droid_jointpos (exterior camera 1 + wrist camera, 7 joint positions + gripper, 15-step chunks of absolute joint targets + gripper). Training data: DecentVLA/droid8_openpi (38,979 DROID episodes from the 8 largest buildings).

Path	What
openpi/	official openpi checkpoint (params/ + assets/droid/norm_stats.json) — serve with plain openpi
server_round_0400.pt	decent-vla checkpoint (resume / continue training) + openpi_pi05_droid8_cent_local.yaml, norm_stats.json, training log
droid_jointpos_main.py	real-robot client for a DROID platform (joint-position control)

(Files arrive as a pull request from the training run; see Community → Pull requests until it is merged.)

Evaluate anywhere — RoboLab or a real DROID robot

Any machine with an NVIDIA GPU (RoboLab: RTX, >= 48 GB; serving only: >= 16 GB).

huggingface-cli download DecentVLA/droid8_pi05_cent --include "openpi/*" --local-dir droid8_cent
# RoboLab's own openpi server (xuningy/openpi fork, installed per NVLabs/RoboLab policies/pi0_family/README.md):
XLA_PYTHON_CLIENT_MEM_FRACTION=0.5 uv run scripts/serve_policy.py --port 8000 policy:checkpoint \
    --policy.config=pi05_droid_jointpos --policy.dir=$(realpath droid8_cent/openpi)

• RoboLab: uv run python policies/pi0_family/run.py --policy pi05 --headless --tag color --remote-host <server> --remote-port 8000
• Real DROID robot: python droid_jointpos_main.py --remote-host <server> --remote-port 8000 --external-camera left ... (joint-position control; per-step joint safety limit --max-joint-step, default 0.25 rad).

Runbook and code: decent-vla examples/droid/README_droid8.md (tag droid8-v1).