adipotnis/m2-base-embzero-tilt-cf-robowarp

M2-Base — Tilted-Flow Source from π₀.₅ Base (RoboWarp)

The tilt method (source-space counterfactual coupling) transplanted to the base family: π₀.₅ base finetuned on LIBERO-Spatial with the flow's source tilted toward the frozen base model's own emb-zero counterfactual action chunks. No pretrained LIBERO model is used anywhere. Trained with openpi (JAX) on 2× NVIDIA GH200.

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

Method

a_prior = pi05_base(o, emb-zero mask), 10-step sample     precomputed per frame, frozen
x₁ = α·a_prior + √(1−α²)·ε     w.p. p     (else ε ~ N(0, I))
u  = x₁ − a = α(a_prior − a) + √(1−α²)·ε
L(θ) = ‖v_θ(x_t, t | o) − u‖²

The regressed velocity on tilted samples is the correction from the LIBERO-naive, evidence-blind behavior to the demonstration. Counterfactual = emb-zero: the base camera's top-attended SigLIP patch tokens zeroed post-encoding (robowarp layer-12 masks, top-15 % pixels, block-pooled to the 16×16 patch grid at ≥ 50 % coverage). Inference unchanged: single pass from N(0, I).

Hyperparameters

Init = prior drafter	gs://openpi-assets/checkpoints/pi05_base (frozen for priors)
Model	π₀.₅, action_horizon 10, state-blind
Data	LIBERO-Spatial only: 432/1693 episodes, 52,970 frames
Coupling	α = 0.6, p = 0.5
Priors	52,970 chunks, 10 Euler steps, emb-zero sampling, seed 7000
Batch / steps	112 / 2,300 (≈4.9 epochs — the base-family recipe, matching M4/M5/M6/M7/baseline)
LR / optimizer	3.3e-5 → 3.3e-6 cosine, warmup 200 · AdamW, clip 1.0, EMA 0.999

Caveats

• Training loss not comparable to other arms (tilted sources shrink targets). Sim eval only.
• ~4.9 epochs from base. α, p not swept.
• Family: pi05-base-spatial-fft-baseline (matched vanilla twin), M4R/M5R (steering), m6-contrastive-cf-robowarp, m7-guidance-delta-cf-robowarp. Released-family twin of the method: adipotnis/m2-train-time-cf-robowarp.