Modelpublic

tencent/Hy-Embodied-RxBrain-1.0

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

57likes315downloads

Model Card

<div align="center"> <h1>RxBrain</h1> <p><b>Embodied Cognition Foundation Model with Joint Language–Visual Reasoning and Imagination</b></p> <p><i>Tencent Robotics X × Futian Laboratory × Tencent Hy Team</i></p>

</div>

🔥 Updates

—`[2026-07]` 🎉 We release Hy-Embodied-RxBrain-1.0 — the technical report, official inference code, and model weights.

📖 Introduction

RxBrain (Hy-Embodied-RxBrain-1.0) is a unified multimodal foundation model for embodied cognition — a single model that couples language reasoning with visual imagination to deliver three core capabilities:

—🤖 Embodied Understanding & Reasoning — question answering and chain-of-thought over images and multi-frame video.
—🔮 World State Prediction — imagine the near-future frames an action produces in the physical world.
—🧩 Joint Subgoal Planning — decompose a task into steps, emitting for each step both the next action (language) and the goal image it should reach (vision).

These capabilities are unified through interleaved generation: within a single autoregressive sequence RxBrain alternates reasoning text and flow-matched imagined frames — a learned <Image> token decides when to imagine — so an embodied plan couples what to do with what the world should look like, step by step.

⭐️ Key Features

—🧠 Unified Mixture-of-Transformers (MoT): A ~6.2B-parameter backbone with modality-specific pathways (text / vision / generation), so understanding and image synthesis share one autoregressive model instead of separate towers.
—🎨 Flow-Matching Image Head: Imagined frames are produced by a flow-matching head decoding into a frozen FLUX VAE latent space, enabling text-to-image, multi-frame world-model rollout, and goal-image planning.
—🔗 Interleaved Reasoning + Imagination: Text reasoning and generated frames are emitted in one sequence, coupling symbolic plans with visual goals.

📋 TODO

—[ ] Open-source RxBrain-Bench, a comprehensive benchmark for embodied cognition.
—[ ] Open-source the fine-tuning code for Visual Question Answering (VQA), Multi Frames Generation, and Interleave Generation.
—[x] Release the technical report, inference code, and model weights.

🛠️ Dependencies and Installation

Prerequisites

—🖥️ Operating System: Linux (recommended)
—🐍 Python: 3.10+
—⚡ CUDA: 12.x, an NVIDIA GPU (required for flash-attn)
—🔥 PyTorch: 2.10

Installation

1.Install the specific Transformers version required for this model (it provides the hunyuan_vl_mot backbone that unified_mot builds on):

bash

pip install git+https://github.com/huggingface/transformers@9293856c419762ebf98fbe2bd9440f9ce7069f1a

Note: A stock transformers release does not yet include hunyuan_vl_mot; this pinned commit is required. We will merge the improvements into the Transformers main branch later.

1.Clone the inference code and install the remaining dependencies:

bash

git clone https://github.com/Tencent-Hunyuan/Hy-Embodied-RxBrain-1.0.git
cd Hy-Embodied-RxBrain-1.0
pip install -r requirements.txt

Model Download

Component	Params	Source
Hy-Embodied-RxBrain-1.0	~6.2 B	🤗 tencent/Hy-Embodied-RxBrain-1.0
FLUX VAE (`ae.safetensors`)	83.8 M	Obtain from the FLUX distribution

Download the weights to a local directory — the loader reads the checkpoint files directly, so --ckpt must be a local path, not the Hub repo id:

bash

pip install -U "huggingface_hub[cli]"
hf download tencent/Hy-Embodied-RxBrain-1.0 --local-dir ./Hy-Embodied-RxBrain-1.0

The VQA (understanding) path needs only the main weights. Image generation (T2I / world-model rollout / interleaved planning) additionally requires the external FLUX VAE ae.safetensors.

🚀 Quick Start with Transformers

Load the Transformers processor together with the UnifiedMoT classes shipped in this repo, then run understanding (VQA). Run this from the repo root so the model package is importable, and point MODEL_PATH at your local download (see Model Download).

python

import torch
from transformers.models.hunyuan_vl_mot import HunYuanVLMoTProcessor
from model import UnifiedMoTForConditionalGeneration, maybe_init_generation_path
from vqa_inference import answer

MODEL_PATH = "./Hy-Embodied-RxBrain-1.0"  # local checkpoint directory, not the Hub id
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
dtype = torch.bfloat16

# Load processor & model
processor = HunYuanVLMoTProcessor.from_pretrained(MODEL_PATH, trust_remote_code=True)
model = UnifiedMoTForConditionalGeneration.from_pretrained(MODEL_PATH, dtype=dtype)
maybe_init_generation_path(model, model_load_path=MODEL_PATH)  # wires up the generation path
model.to(device).eval()

# Ask a question about an image
text = answer(
    model, processor,
    image_paths=["demo_cases/bridgev2_move_toy/input/obs_1.jpg"],
    question="What objects are on the stovetop, and where is the green toy?",
    device=device, dtype=dtype, max_new_tokens=256,
)
print(text)

Note: RxBrain uses a custom interleaved text/image decoding loop rather than the standard model.generate API. The answer(...) helper (in vqa_inference.py) wraps that loop for the understanding case; image generation and planning have their own entry points below.

The same tasks are also available as ready-to-run scripts:

<details open> <summary><b>① Visual Question Answering (VQA)</b> — image(s) + question → answer text</summary>

Pure autoregressive text understanding — no VAE / flow-matching needed.

bash

python vqa_inference.py \
    --ckpt ./Hy-Embodied-RxBrain-1.0 \
    --images demo_cases/bridgev2_move_toy/input/obs_1.jpg \
    --question "What objects are on the stovetop, and where is the green toy?" \
    --max_new_tokens 256

</details>

<details open> <summary><b>② Text-to-Image (T2I)</b></summary>

bash

python text2image_inference.py \
    --ckpt ./Hy-Embodied-RxBrain-1.0 --vae /path/to/ae.safetensors \
    --prompt "a watercolor painting of a cat" \
    --height 256 --width 256 --num_steps 25 --out out.png

# with classifier-free guidance
python text2image_inference.py \
    --ckpt ./Hy-Embodied-RxBrain-1.0 --vae /path/to/ae.safetensors \
    --prompt "a watercolor painting of a cat" \
    --cfg_scale 5.0 --num_steps 50 --out out.png

</details>

<details open> <summary><b>③ Multi-Frame World-Model Rollout</b> — imagine future frames from an observation</summary>

bash

python multiframe_inference.py \
    --ckpt ./Hy-Embodied-RxBrain-1.0 --vae /path/to/ae.safetensors \
    --frames /path/to/obs.jpg --task "imagine the next frames" \
    --num_frames 4 --num_steps 50 --out_dir multiframe_out

</details>

<details> <summary><b>④ Interleaved Embodied Planning</b> — text plan + goal images, step by step</summary>

Runs interleaved planning on a bundled scene. See `demo_cases/README.md` for details.

bash

CASE=umi_fold_sock
python interleave_inference.py \
    --ckpt ./Hy-Embodied-RxBrain-1.0 --vae /path/to/ae.safetensors \
    --frames  demo_cases/$CASE/input/*.jpg \
    --task    "$(cat demo_cases/$CASE/prompt.txt)" \
    --max_frames 5 --num_steps 50 --out_dir out_$CASE

</details>

📊 Evaluation

RxBrain is evaluated on embodied understanding, spatial reasoning, and imagination/generation benchmarks. For detailed metrics and methodology, please refer to our technical report.