| { |
| "schema": "parallelvla.factorized_router_vs_skillvla_r4.v1", |
| "conclusion": "The successful r4 policy is teacher-driven; the per-arm student was not the deployed action source.", |
| "skillvla_r4": { |
| "initialization": "adapted full-vision all-actions step-2000 joint anchor", |
| "trainable": "router/semantic/student auxiliaries only", |
| "deployed_action_source": "joint teacher because student_mix=0.0", |
| "inactive_action_weight": 1, |
| "cadence": { |
| "async": 50, |
| "sync": 10, |
| "threshold": 0.53, |
| "confirmations": 2, |
| "boundary": 20 |
| }, |
| "data": { |
| "episodes": 50, |
| "frames": 10414, |
| "distribution": "canonical dynamic-current-frame" |
| }, |
| "transplant": { |
| "recipient_step": 500, |
| "donor_step": 500, |
| "parameters": [ |
| "action_out_proj/kernel", |
| "action_out_proj/bias" |
| ], |
| "donor_weight": 0.6875 |
| }, |
| "fixed100": { |
| "suite": "parallelvla_eval_pcd", |
| "success": 57, |
| "collision": 8, |
| "matched_control_success": 55 |
| } |
| }, |
| "factorized_router": { |
| "initialization": "raw PI0.5 base", |
| "trainable": "full 3.35B backbone", |
| "inactive_arm_loss": "excluded static-arm action dimensions", |
| "cadence": "server returned 50 actions for every request; failed rollouts were 14x50=700", |
| "router": "independent left/right stage prediction with observed reach/hold oscillation", |
| "data": { |
| "episodes": 100, |
| "rows": 21645, |
| "distribution": "50 canonical plus 50 random-delay" |
| } |
| }, |
| "controlled_reproduction": { |
| "principle": "restore the known-successful training and deployment chain before changing one factor", |
| "max_updates": 2000, |
| "first_native_gate": 500, |
| "continue_condition": ">=1/5 real native fixed-seed successes" |
| } |
| } |
|
|
