Global L2-SP target manipulation pilot: view_large / retention
This source-free experiment uses 1000 native IDM updates, B32, and raw non-EMA step999. Its deployed split74 contains native43 FP32, frozen8 BF16, source-wrist23 BF16, and the original normalization assets. Inference restores BF16.
For the retention arm the objective is L_IDM + lambda * R, with R = 0.5 * SUM over all native43 parameter elements (theta - stopgrad(theta0))^2. The one external lambda is 0.001; the common planned coefficient for view_large and yaw30 is 0.001. There is no module, leaf or parameter-count normalization. The fixed arm uses lambda0 and the native IDM optimizer route. Source43 anchors are independent frozen FP32 copies. No teacher policy forward, source-output distillation, flow anchors, or source observations contribute to this penalty. Analytic gradients are added before native clipping and AdamW; this is not equivalent to module learning-rate changes and does not guarantee smaller gradient norms or target manipulation retention.
The first coefficient is a predeclared heuristic informed only by completed trajectory endpoint parameter scales, not actual same-state calibration or source/target rollout selection. Its units differ from output-velocity MSE lambda. The exact frozen coefficient specification is included with this package.
Actual initialization used the legacy local native51 source. The public model Donghyun1228/pi05-libero-main-policy-retention-source-split74-20261013 at 24268da2aba5f690c879788d98e3455a2a579d8f is a canonical-equivalent split74 deployment reference, not the actual training input. Target interaction dataset: Donghyun1228/libero-dense-contact-sweep-d2-20261002 at 9f8f600ce0e8f5df51c0e847c65a1e09f29d8494. Use pi05_libero_source_free_split_vision_idm with inference revision 0f6471d99dfc81fabe7c3ab65c141d7640b9ff08.
The primary outcome is target manipulation success: 40 LIBERO tasks with 10 episodes each, environment seed 7, 8 clients/GPU and MuJoCo 3.2.3. Yaw uses both established compensation and action-rotation flags. Policy noise follows shared server streams, not episodewise paired streams. Original-environment success is not a required outcome and is not substituted for target success. See evaluation_target400.json and evaluation_protocols.json for actual evidence.
This single pilot does not establish efficacy or an optimal coefficient.