OriginX

OriginX is a robot policy for RoboCasa365, developed by Qinzhen Ma using Xiaomi-Robotics-1-RoboCasa365 as its foundation.

Project website · Code and inference guide · Technical report

Complete model weights

This repository contains the complete inference weight set: the three full base-model shards, tokenizer and processor assets, the A2000 Action LoRA adapter, and the B2000 continuous conditioning branch. The base-model weights are included here; a separate download from another model repository is not required.

Component	Files
Full frozen base	model-00001-of-00003.safetensors, model-00002-of-00003.safetensors, model-00003-of-00003.safetensors
A2000 Action LoRA	adapter-originx-2000.pt
B2000 conditioning branch	branch-00002000.pt
Model assets	Configuration, weight index, tokenizer, and processor files

The base has 5,053,149,696 parameters stored in BF16. Its three shards total 10,106,433,336 bytes. The adaptation files are the unchanged weights used in the recorded evaluation; they remain separate to preserve the original loading and numerical behavior.

Release naming. A2000 is the public name of the existing adapter. It completed 1,613 optimizer updates over 80,000 sampled windows; B2000 completed 2,000 updates over 256,000 sampled windows. Renaming does not alter weight bytes, training history, or evaluation results.

Run the model

Implementation code lives on GitHub. Clone that repository and follow PORTABLE_INFERENCE.md. The prepare_originx.py helper downloads this model snapshot, verifies the original asset hashes, and combines its weights with the three matching model implementation files from GitHub. It prepares assets/base and assets/weights for the loader. No Python source files are hosted in this Hugging Face repository.

The frozen upstream revision is 3a6d0293bfa90759d34a7fc48c2c62413cd7bcf4. SHA256SUMS covers the published model assets. The publication loader has CPU contract checks; a new real-model GPU parity run through that helper has not been performed.

Evaluation

The complete author-run native-reset evaluation recorded 1,496/2,500 successes (59.84%) across 50 tasks, with 50 episodes per task. All 1,004 policy failures are retained, with zero missing or infrastructure-unknown outcomes.

Split	Successes / episodes	Rate
Atomic-Seen	747 / 900	83.00%
Composite-Seen	479 / 800	59.875%
Composite-Unseen	270 / 800	33.75%

RoboCasa submission #26 is pending organizer review. An earlier separate 600-pair comparison found a 1.00 percentage-point difference with reported 95% interval [-2.17, 4.17] and McNemar p=0.6173; it did not establish a reliable gain. The full evaluation has no matched full baseline or branch ablation. The technical report retains the protocol, training lineage, and limitations; evaluation evidence remains on GitHub.

License

Apache-2.0. See LICENSE and NOTICE for attribution. Source code, reports, and evaluation archives are maintained on GitHub; this repository distributes model weights and their configuration/documentation.