M6 — Contrastive Counterfactual Velocity (RoboWarp)
π₀.₅ base finetuned on LIBERO-Spatial with a contrastive counterfactual objective: the CFM regression target stays plain (no steering), and an InfoNCE term makes the student's predicted velocity rank its own target above the evidence-blind counterfactual velocity. Teacher = the run's own EMA of the student (BYOL-style, τ = 0.999); counterfactual = the factual image with the top-attended SigLIP patch tokens zeroed post-encoding (emb-zero). Trained with openpi (JAX) on 2× GH200.
Contents: params/ (Orbax weights) + assets/ (spatial-suite quantile normalization
stats). No train_state/, so this cannot be resumed.
Method (M6)
eps ~ N(0, I), t ~ Beta(1.5, 1)·0.999 + 0.001
x_t = t·eps + (1−t)·a u = eps − a
v_neg = v_ξ(x_t, t | o, emb-zero mask) EMA target net ξ, stop-grad
z = v_θ(x_t, t | o) student velocity (anchor, with gradient)
L = ‖z − u‖² + λ·InfoNCE(z; positive u_i, negatives {v_neg,i} ∪ {u_j}_{j≠i})
(cosine similarity, temperature τ_nce; every sample, no gating)
every step: ξ ← 0.999·ξ + 0.001·θ
Unlike the steering family (M3–M5, M7), the repulsion is self-normalizing: the InfoNCE gradient is strong only while the student's velocity is confusable with the evidence-blind one and fades once separated — no constant-ω target distortion, no ω to tune. In-batch negatives come free. Inference unchanged: single pass from N(0, I).
Hyperparameters
| Init | gs://openpi-assets/checkpoints/pi05_base |
| Teacher | EMA of the student, τ = 0.999 (soft-updated every step) |
| Model | π₀.₅, action_horizon 10, state-blind |
| Data | LIBERO-Spatial only: 432/1693 episodes, 52,970 frames |
| Counterfactual | emb-zero: 16×16 SigLIP patch grid, patch zeroed at ≥ 50 % attended pixels (masks: robowarp layer-12, top-15 %, from released pi05_libero) |
| Contrastive | λ = 0.02, τ_nce = 0.1, hard negative + in-batch negatives |
| Batch / steps | 112 / 2,300 (≈4.9 epochs) |
| 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; evaluate by sim success (LIBERO-PRO spatial/swap).
- λ and τ_nce not swept. ~4.9 epochs from base.
- Family:
pi05-base-spatial-fft-baseline(matched vanilla twin), M4/M5 (steering siblings),m7-guidance-delta-cf-robowarp(delta-distillation sibling).