farm_uf850_home_fft

farm_uf850_home_fft

Full fine-tune (every parameter trainable, no LoRA/GSE) of pi0.5 on NoahWeiss/farm_uf850_home_full (revision v2.0): 1153 teleoperated UFactory UF850 episodes across 20 home manipulation tasks, recorded with a Quest 3 rig.

Held-out split

63 episodes (2-5 per task, stratified, seed 153) were excluded from training and reserved for evaluation. The full labeled list is in heldout_home_full.json in this repo. Training used the remaining 1090 episodes (~377k frames). Norm stats were computed on the train split only.

Recipe

• openpi pi05 config pi05_farm_home_fft: absolute joint actions (use_delta_joint_actions=False), action_horizon=10, EMA 0.999 (saved params are the EMA copy)
• 8x NVIDIA H200, fsdp_devices=2 (4 data-parallel replicas), batch 64
• AdamW, cosine schedule: 2k warmup, peak LR 3.5e-5, decay to 3.5e-6
• heavy image domain randomization (brightness/contrast/saturation/hue, per-channel gamma, grayscale, blur, crop/rotate)
• ~12 h single-node run; checkpoints streamed here every save_interval steps as step-<N> tags (params + assets, inference-ready)

Training/eval code: the FARM repo (model/amp_fft/).