adipotnis/pi05-base-spatial-fft-baseline

π₀.₅-base LIBERO-Spatial FFT — baseline arm for M4

Plain full finetune of π₀.₅ base (pi05_base, no prior LIBERO training) on the LIBERO-Spatial suite only, with standard conditional flow matching — no counterfactual machinery. This is the matched baseline for adipotnis/m4-bootstrap-cfsteer-robowarp: identical init, data, normalization, state-blindness, batch, LR schedule and step count — the only difference is the vanilla loss (no bootstrapped teacher, no target steering). Trained with openpi (JAX) on 2× NVIDIA GH200.

Contents: params/ (Orbax weights) + assets/ (spatial-suite quantile normalization stats — required at inference). No train_state/, so this cannot be resumed.

Hyperparameters

Init	gs://openpi-assets/checkpoints/pi05_base
Model	π₀.₅ (pi05=True, action_horizon 10, discrete_state_input=False — state-blind)
Data	physical-intelligence/libero v2.0, LIBERO-Spatial suite only: 432/1693 episodes, 52,970 frames (task indices 30–39)
Normalization	spatial-suite quantile stats computed from the training data (same file as the M4 run)
Loss	standard CFM: ‖v_θ(x_t, t) − (ε − a)‖², ε ~ N(0, I)
Batch / steps	112 (2× GH200, FSDP — matched to M4, though no teacher is resident) / 2,300 (≈4.9 epochs)
LR	3.3e-5 → 3.3e-6 cosine, warmup 200 (matched to M4)
Optimizer	AdamW β=(0.9, 0.95), eps 1e-8, wd 1e-10, grad-clip 1.0, EMA 0.999
Precision	bfloat16

Caveats

• Not evaluated in sim at upload time. ~4.9 epochs from BASE (vs ~48 in the reference recipe).
• Loss IS comparable to nothing here either — different init/norm-stats lineage than the pi05_libero-init arms. Evaluate by simulator success rate (LIBERO-PRO spatial/swap).
• Comparison twin: adipotnis/m4-bootstrap-cfsteer-robowarp (identical run + bootstrapped counterfactual steering, ω=0.5 p=0.5 N=1000). Related: adipotnis/pi05-libero-spatial-run1 (also from pi05_base but state-tokenized, batch 160).