farm_uf850_pi05_clean โ ฯ0.5 PiSSA fine-tunes for the FARM UF850 (home_full)
Four servable openpi checkpoints trained 2026-07-19 on NoahWeiss/farm_uf850_home_full (1153 episodes / 397k frames / 30 fps, UF850 arm, base + wrist RealSense):
| Path | What |
|---|---|
farm_soup_s42/14999 |
recommended โ temporal soup (equal-mean of steps 10000+14999), seed 42 |
farm_soup_s43/14999 |
seed-43 twin soup |
farm_clean_s42/14999 |
seed-42 final (EMA 0.999 weights) |
farm_clean_s43/14999 |
seed-43 final (EMA 0.999 weights) |
Recipe: ฯ0.5 base, PiSSA-initialized LoRA r64, masked loss + zero-padded noise
source on the 7 real action dims (of 32), M-draw 4, EMA 0.999, 15k steps,
batch 32. Each dir holds openpi-format params/ + assets/ (norm stats ride
along; optimizer state stripped).
Serving (openpi)
OPENPI_LORA_RANK=64 OPENPI_ZEROPAD=7 \
uv run scripts/serve_policy.py --port 8000 policy:checkpoint \
--policy.config pi05_farm_clean \
--policy.dir <download>/farm_soup_s42/14999
Contract (must match โ details + full arm runbook in
NoahWeiss123/farm model/FARM_CLEAN_ARM_RUNBOOK.md):
OPENPI_LORA_RANK=64 and OPENPI_ZEROPAD=7 at serve; action_horizon=50;
outputs are ABSOLUTE joint targets (6 joints rad + gripper 0..1), 7 dims;
obs keys observation/image (base cam), observation/wrist_image,
observation/state (7,), prompt = a trained task string verbatim;
224x224 resize-with-pad. Client: model/eval_pi05.py with --no-rtc.
The pi05_farm_clean TrainConfig lives in the farm repo cluster patches
(clean-contract port of pi05_farm_pissa with EMA + home_full).