MXI RoboDojo evaluation assets
Policy: mxi_Hybrid_v0_preview; Native Binding s4000 executor and s4250 planner.
Download the complete pinned package using the instructions at:
https://github.com/D-URing/MXI/blob/main/docs/EVALUATOR_HANDOFF.md
Use model/ as the checkpoint, not a single action file. Keep the official
XPolicyLab framework; install only our policy adapter from the code repository.
The package contains the weights, matching normalization and offline runtime,
checksum/install helpers, and original local acceptance results/videos.
Tested: arx_x5/absolute EE16, two96GB policy GPUs, CUDA13-compatible driver.
ACCEPTANCE.json and evidence/ describe bounded local validation, not a full
benchmark score. Those results predate this ActionAdapter runtime update.
ActionAdapter loads with the Action expert from model/base/action.pt and
computes a conditioned residual after action block6. The existing text encoder
encodes the official task identifier; all inputs use the same module forward.
Initialization preserves prior behavior on the checked task set; no new
training or physical rollout is claimed. Module-level checks are recorded in
evidence/action_adapter_validation.json. The full checkpoint preserves
every pre-existing tensor; evidence/checkpoint_tensor_inventory.json
records their shapes, dtypes and hashes. All other weight files, normalization
and inference settings are retained unchanged.