AIRO-Doffy-WRM-Grasp-consensus-two-contact-mlp-fixed-equal-lambda01

Fixed-equal two-expert WRM policy-consensus Diffusion Policy checkpoint.

• Experts: exactly two jointly trained experts, BEAVER (both rings) and vision.
• BEAVER encoder: contact_mlp, using the configured contact representation.
• Fusion: fixed_equal uses invariant weights 0.5/0.5 for both training and inference. Each denoising step fuses the two epsilon predictions as 0.5 * epsilon_beaver + 0.5 * epsilon_vis.
• Expert auxiliary loss weight (lambda_expert): 0.1.
• Seed: 42; batch size: 32; final training step: 100000.
• Artifacts: checkpoints/ contains the numbered consensus checkpoints, last.pt, the resolved configuration, and metrics. Source and provenance files are included only when present in the explicit allowlist.
• SHA-256 of checkpoints/last.pt: d72ba09c1e6e549571cf520c3f7db0d882ed9841463c5533b2ed627153fdb0d1.

The run used online Weights & Biases (W&B) tracking. W&B run data and credentials are excluded from this repository.