weaver-marlowe-robolab-full-left-rm-s0

WEAVER fine-tuned on RoboLab with REAL RoboMeter reward/critic (Marlowe) — model card

Fine-tune of the released WEAVER action-conditioned multi-view world model on Pi0.5 RoboLab demos, trained on Stanford's Marlowe H100 cluster (plan 011, P9). Unlike the sibling model weaver-marlowe-robolab-full-left-s0, whose reward/critic heads were trained on a placeholder success-gated ramp, this model's reward and critic heads are trained on real per-frame RoboMeter progress labels — making them usable for value-based / best-of-N planning.

Data

• 1,045 train + 55 val episodes, 29 RoboLab tasks (the 92 non-empty raw files of the 178-file collection). SD3 latents + CLIP text features, --ext1 left (over_shoulder_left in panel 0, aligned with the Pi0 policy). Frozen val split.
• Reward channel = RoboMeter-4B. Every episode's reward_progress was relabelled by serving Robometer-4B and scoring the exterior (panel-0) view; the original placeholder ramp is retained as reward_placeholder. Verified non-linear (a real signal, not a clock) on all 1,100 episodes.

Training

• 4× H100 DDP (torchrun), global batch 24 (per-GPU 6), 16,000 steps, LR 2e-5, compile off, seed 0, fresh from base WEAVER (exp_name=robolab_full_left_rm_s0). ~2.6 h wall at ~0.56 s/step.

Final validation (step 16,000)

Video-prediction quality tracks the placeholder-reward sibling within the ~±40 single-seed FVD noise band (that model: exterior 136.4 / wrist 383.5) — expected, since the Flow loss is identical and only the reward targets changed. Final reward-head loss 0.005 and critic loss 0.28 (vs the placeholder model's much larger values), i.e. the heads fit the RoboMeter targets tightly.

Reward/critic heads: usable for planning (with a caveat)

These heads learned a real progress signal, not a clock — the intended improvement. Caveat (plan 010): RoboMeter itself was measured as a weaker within-task success discriminator (AUC 0.64) than the DROID-pretrained base heads (0.81). Whether RoboMeter-distilled heads beat the frozen base heads for planning is an open question — evaluate against that baseline (weaver_reward_probe.py) before relying on them.