# xiaomirobotics/xiaomi-robotics-u0 architecture

> Architecture diagram of the GitHub repository xiaomirobotics/xiaomi-robotics-u0, generated by GitDiagram from its file tree, README and a sample of source files (last updated 2026-10-07). It is an AI-made overview: check details against the source.

- Interactive diagram: https://gitdiagram.com/xiaomirobotics/xiaomi-robotics-u0
- Repository: https://github.com/xiaomirobotics/xiaomi-robotics-u0

## Overview

Xiaomi-Robotics-U0 is a multimodal world foundation model with inference and distributed training code. Inference composes task, runtime, and profile settings, then runs AR or FlashAR generation; documented tasks cover T2I, X2I, scene generation, transfer, and interleaved text/video. FlashAR also offers a vLLM batching path. Training uses FSDP, packed data, and Ulysses sequence parallelism. Several inference entry points and training subsystems are represented from the README and tree; their internal wiring is not fully sampled.

Main entry points: [`inference/scripts`](https://github.com/xiaomirobotics/xiaomi-robotics-u0/tree/main/inference/scripts) (Inference Entrypoints), [`training/src/wm_fsdp/train/train.py`](https://github.com/xiaomirobotics/xiaomi-robotics-u0/blob/main/training/src/wm_fsdp/train/train.py) (Training Loop), [`inference/xr_u0_flashar/model.py`](https://github.com/xiaomirobotics/xiaomi-robotics-u0/blob/main/inference/xr_u0_flashar/model.py) (FlashAR Model), [`training/src/wm_fsdp/models/modeling_unis.py`](https://github.com/xiaomirobotics/xiaomi-robotics-u0/blob/main/training/src/wm_fsdp/models/modeling_unis.py) (Training Model)

## Diagram (Mermaid)

```mermaid
flowchart TD

subgraph group_inference["Inference"]
  node_config["Config Composition<br/>[compose.py]"]
  node_taskconfigs["Task Configs"]
  node_inferentry["Inference Entrypoints"]
end

subgraph group_ar["AR Model"]
  node_arengine["AR Generation<br/>[generation.py]"]
  node_seqengine["Sequence Generation"]
  node_unismodel["UNIS Model<br/>[modeling_unis.py]"]
  node_uniscfg["UNIS Config"]
  node_visiontok["Vision Tokenizer<br/>[ibq.py]"]
  node_depth["Depth Preprocessing"]
end

subgraph group_flashar["FlashAR"]
  node_flashmodel["FlashAR Model<br/>[model.py]"]
  node_flashoutputs["FlashAR Outputs<br/>[outputs.py]"]
  node_vllmapi["vLLM API<br/>[api.py]"]
  node_vllmrun["vLLM Runtime<br/>[runtime.py]"]
end

subgraph group_training["Training"]
  node_trainingdata["Training Data"]
  node_trainentry["Training Loop<br/>[train.py]"]
  node_trainmodel["Training Model<br/>[modeling_unis.py]"]
  node_trainconfig["Training Config"]
  node_traintokenizer["Training Tokenizers"]
  node_fsdp["FSDP Utilities<br/>[fsdp_utils.py]"]
  node_ulysses["Ulysses Parallelism<br/>[ulysses.py]"]
end

node_user(("User"))
node_weights[("Model Weights")]
node_result["Generated Result"]

node_user -->|"invokes"| node_inferentry
node_config -->|"loads task"| node_taskconfigs
node_inferentry -.->|"uses settings"| node_config
node_inferentry -.->|"runs AR"| node_arengine
node_inferentry -.->|"runs FlashAR"| node_flashmodel
node_arengine -.->|"generates with"| node_unismodel
node_unismodel -->|"uses config"| node_uniscfg
node_visiontok -.->|"loads tokenizer weights"| node_weights
node_inferentry -->|"preprocesses depth"| node_depth
node_flashmodel -->|"wraps backbone"| node_unismodel
node_flashmodel -.->|"produces outputs"| node_flashoutputs
node_vllmapi -.->|"dispatches runtime"| node_vllmrun
node_user -.->|"requests batched generation"| node_vllmapi
node_vllmrun -->|"uses FlashAR model"| node_flashmodel
node_inferentry -.->|"returns generation"| node_result
node_user -->|"starts training"| node_trainentry
node_trainentry -->|"loads batches"| node_trainingdata
node_trainentry -->|"trains model"| node_trainmodel
node_trainmodel -->|"uses config"| node_trainconfig
node_trainentry -->|"uses utilities"| node_fsdp
node_trainentry -->|"uses parallelism"| node_ulysses
node_trainmodel -.->|"uses tokenizer"| node_traintokenizer
node_trainentry -.->|"saves checkpoints"| node_weights

click node_config "https://github.com/xiaomirobotics/xiaomi-robotics-u0/blob/main/inference/configs/compose.py"
click node_taskconfigs "https://github.com/xiaomirobotics/xiaomi-robotics-u0/tree/main/inference/configs/tasks"
click node_inferentry "https://github.com/xiaomirobotics/xiaomi-robotics-u0/tree/main/inference/scripts"
click node_arengine "https://github.com/xiaomirobotics/xiaomi-robotics-u0/blob/main/inference/xr_u0_ar/generation.py"
click node_seqengine "https://github.com/xiaomirobotics/xiaomi-robotics-u0/blob/main/inference/xr_u0_ar/sequence_generation.py"
click node_unismodel "https://github.com/xiaomirobotics/xiaomi-robotics-u0/blob/main/inference/xr_u0_ar/modeling_unis.py"
click node_uniscfg "https://github.com/xiaomirobotics/xiaomi-robotics-u0/blob/main/inference/xr_u0_ar/configuration_unis.py"
click node_visiontok "https://github.com/xiaomirobotics/xiaomi-robotics-u0/blob/main/inference/xr_u0_ar/vision_tokenizer/ibq.py"
click node_depth "https://github.com/xiaomirobotics/xiaomi-robotics-u0/blob/main/inference/xr_u0_ar/depth_preprocess.py"
click node_flashmodel "https://github.com/xiaomirobotics/xiaomi-robotics-u0/blob/main/inference/xr_u0_flashar/model.py"
click node_flashoutputs "https://github.com/xiaomirobotics/xiaomi-robotics-u0/blob/main/inference/xr_u0_flashar/outputs.py"
click node_vllmapi "https://github.com/xiaomirobotics/xiaomi-robotics-u0/blob/main/inference/xr_u0_flashar/vllm/api.py"
click node_vllmrun "https://github.com/xiaomirobotics/xiaomi-robotics-u0/blob/main/inference/xr_u0_flashar/vllm/runtime.py"
click node_trainingdata "https://github.com/xiaomirobotics/xiaomi-robotics-u0/tree/main/training/src/wm_fsdp/data"
click node_trainentry "https://github.com/xiaomirobotics/xiaomi-robotics-u0/blob/main/training/src/wm_fsdp/train/train.py"
click node_trainmodel "https://github.com/xiaomirobotics/xiaomi-robotics-u0/blob/main/training/src/wm_fsdp/models/modeling_unis.py"
click node_trainconfig "https://github.com/xiaomirobotics/xiaomi-robotics-u0/blob/main/training/src/wm_fsdp/models/configuration_unis.py"
click node_traintokenizer "https://github.com/xiaomirobotics/xiaomi-robotics-u0/tree/main/training/src/wm_fsdp/tokenizers"
click node_fsdp "https://github.com/xiaomirobotics/xiaomi-robotics-u0/blob/main/training/src/wm_fsdp/train/fsdp_utils.py"
click node_ulysses "https://github.com/xiaomirobotics/xiaomi-robotics-u0/blob/main/training/src/wm_fsdp/train/ulysses.py"

classDef toneNeutral fill:#f8fafc,stroke:#334155,stroke-width:1.5px,color:#0f172a
classDef toneBlue fill:#dbeafe,stroke:#2563eb,stroke-width:1.5px,color:#172554
classDef toneAmber fill:#fef3c7,stroke:#d97706,stroke-width:1.5px,color:#78350f
classDef toneMint fill:#dcfce7,stroke:#16a34a,stroke-width:1.5px,color:#14532d
classDef toneRose fill:#ffe4e6,stroke:#e11d48,stroke-width:1.5px,color:#881337
classDef toneIndigo fill:#e0e7ff,stroke:#4f46e5,stroke-width:1.5px,color:#312e81
classDef toneTeal fill:#ccfbf1,stroke:#0f766e,stroke-width:1.5px,color:#134e4a
class node_config,node_taskconfigs,node_inferentry,node_user toneBlue
class node_arengine,node_seqengine,node_unismodel,node_uniscfg,node_visiontok,node_depth,node_weights toneAmber
class node_flashmodel,node_flashoutputs,node_vllmapi,node_vllmrun toneMint
class node_trainingdata,node_trainentry,node_trainmodel,node_trainconfig,node_traintokenizer,node_fsdp,node_ulysses toneRose
class node_result toneIndigo
```

## Components

### Inference

- [Config Composition](https://github.com/xiaomirobotics/xiaomi-robotics-u0/blob/main/inference/configs/compose.py): `inference/configs/compose.py`
- [Task Configs](https://github.com/xiaomirobotics/xiaomi-robotics-u0/tree/main/inference/configs/tasks): `inference/configs/tasks`
- [Inference Entrypoints](https://github.com/xiaomirobotics/xiaomi-robotics-u0/tree/main/inference/scripts): `inference/scripts`

### AR Model

- [AR Generation](https://github.com/xiaomirobotics/xiaomi-robotics-u0/blob/main/inference/xr_u0_ar/generation.py): `inference/xr_u0_ar/generation.py`
- [Sequence Generation](https://github.com/xiaomirobotics/xiaomi-robotics-u0/blob/main/inference/xr_u0_ar/sequence_generation.py): `inference/xr_u0_ar/sequence_generation.py`
- [UNIS Model](https://github.com/xiaomirobotics/xiaomi-robotics-u0/blob/main/inference/xr_u0_ar/modeling_unis.py): `inference/xr_u0_ar/modeling_unis.py`
- [UNIS Config](https://github.com/xiaomirobotics/xiaomi-robotics-u0/blob/main/inference/xr_u0_ar/configuration_unis.py): `inference/xr_u0_ar/configuration_unis.py`
- [Vision Tokenizer](https://github.com/xiaomirobotics/xiaomi-robotics-u0/blob/main/inference/xr_u0_ar/vision_tokenizer/ibq.py): `inference/xr_u0_ar/vision_tokenizer/ibq.py`
- [Depth Preprocessing](https://github.com/xiaomirobotics/xiaomi-robotics-u0/blob/main/inference/xr_u0_ar/depth_preprocess.py): `inference/xr_u0_ar/depth_preprocess.py`

### FlashAR

- [FlashAR Model](https://github.com/xiaomirobotics/xiaomi-robotics-u0/blob/main/inference/xr_u0_flashar/model.py): `inference/xr_u0_flashar/model.py`
- [FlashAR Outputs](https://github.com/xiaomirobotics/xiaomi-robotics-u0/blob/main/inference/xr_u0_flashar/outputs.py): `inference/xr_u0_flashar/outputs.py`
- [vLLM API](https://github.com/xiaomirobotics/xiaomi-robotics-u0/blob/main/inference/xr_u0_flashar/vllm/api.py): `inference/xr_u0_flashar/vllm/api.py`
- [vLLM Runtime](https://github.com/xiaomirobotics/xiaomi-robotics-u0/blob/main/inference/xr_u0_flashar/vllm/runtime.py): `inference/xr_u0_flashar/vllm/runtime.py`

### Training

- [Training Data](https://github.com/xiaomirobotics/xiaomi-robotics-u0/tree/main/training/src/wm_fsdp/data): `training/src/wm_fsdp/data`
- [Training Loop](https://github.com/xiaomirobotics/xiaomi-robotics-u0/blob/main/training/src/wm_fsdp/train/train.py): `training/src/wm_fsdp/train/train.py`
- [Training Model](https://github.com/xiaomirobotics/xiaomi-robotics-u0/blob/main/training/src/wm_fsdp/models/modeling_unis.py): `training/src/wm_fsdp/models/modeling_unis.py`
- [Training Config](https://github.com/xiaomirobotics/xiaomi-robotics-u0/blob/main/training/src/wm_fsdp/models/configuration_unis.py): `training/src/wm_fsdp/models/configuration_unis.py`
- [Training Tokenizers](https://github.com/xiaomirobotics/xiaomi-robotics-u0/tree/main/training/src/wm_fsdp/tokenizers): `training/src/wm_fsdp/tokenizers`
- [FSDP Utilities](https://github.com/xiaomirobotics/xiaomi-robotics-u0/blob/main/training/src/wm_fsdp/train/fsdp_utils.py): `training/src/wm_fsdp/train/fsdp_utils.py`
- [Ulysses Parallelism](https://github.com/xiaomirobotics/xiaomi-robotics-u0/blob/main/training/src/wm_fsdp/train/ulysses.py): `training/src/wm_fsdp/train/ulysses.py`

### Other components

- **User**
- **Model Weights**
- **Generated Result**

## Connections

- User → Inference Entrypoints: invokes (evidence: `README.md`)
- Config Composition → Task Configs: loads task (evidence: `inference/configs/compose.py`)
- Inference Entrypoints → Config Composition: uses settings
- Inference Entrypoints → AR Generation: runs AR
- Inference Entrypoints → FlashAR Model: runs FlashAR (evidence: `inference/scripts/inference_flashar.py`)
- AR Generation → UNIS Model: generates with
- UNIS Model → UNIS Config: uses config (evidence: `inference/xr_u0_ar/modeling_unis.py`)
- Vision Tokenizer → Model Weights: loads tokenizer weights
- Inference Entrypoints → Depth Preprocessing: preprocesses depth (evidence: `inference/scripts/inference_flashar.py`)
- FlashAR Model → UNIS Model: wraps backbone (evidence: `inference/xr_u0_flashar/model.py`)
- FlashAR Model → FlashAR Outputs: produces outputs
- vLLM API → vLLM Runtime: dispatches runtime
- User → vLLM API: requests batched generation
- vLLM Runtime → FlashAR Model: uses FlashAR model (evidence: `inference/xr_u0_flashar/vllm/runtime.py`)
- Inference Entrypoints → Generated Result: returns generation
- User → Training Loop: starts training (evidence: `README.md`)
- Training Loop → Training Data: loads batches (evidence: `training/src/wm_fsdp/train/train.py`)
- Training Loop → Training Model: trains model (evidence: `training/src/wm_fsdp/train/train.py`)
- Training Model → Training Config: uses config (evidence: `training/src/wm_fsdp/models/modeling_unis.py`)
- Training Loop → FSDP Utilities: uses utilities (evidence: `training/src/wm_fsdp/train/train.py`)
- Training Loop → Ulysses Parallelism: uses parallelism (evidence: `training/src/wm_fsdp/train/train.py`)
- Training Model → Training Tokenizers: uses tokenizer
- Training Loop → Model Weights: saves checkpoints
