AIRO-Doffy-WRM-Grasp-consensus-two-contact-mlp-detached-router-lambda0

Two jointly trained diffusion experts: BEAVER (both rings) and vision. The unified BEAVER encoder uses 144 strict-<10 mm contact bits plus 9 sensor-enable features. A contact is valid only for status 5 or 9, a finite distance >= 0, and the presence gate. The two native observations are independent inputs to a shared MLP. No GRU and no raw distances are used.

Observation-only routing uses detach(z_B), detach(z_V), and q; x_t and t are not router inputs. The expert auxiliary loss weight is lambda_expert=0.0. Single task: cluster9, 1 GPU, 8 CPUs, batch size 32, online W&B tracking. Training ends at 100000 steps. Artifacts and source are in checkpoints/. last.pt SHA-256: b96f0f6e1e1ab9f762f30da5af564905177bebbf94f59fe08a1420853f8329a7