Main-policy retention experiment: view_large / retention
This source-free adaptation checkpoint contains the raw, non-EMA parameters after 1000 optimizer updates: native43 FP32, frozen8 BF16, and source-wrist23 BF16, with the original normalization assets. The retention arm adds one shared flow velocity consistency coefficient; fixed and scalar arms are comparison controls. The objective is preservation of the main manipulation policy, not improved IDM.
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. Inference restores BF16. Both target and original environment evaluations contain 40 tasks with 10 episodes each, seed 7, 8 clients/GPU, and MuJoCo 3.2.3. Original uses agentview, yaw0, with no compensation or action rotation. Evaluation JSONs preserve results. This individual checkpoint does not establish a causal retention benefit.