---
title: groot-mind2act
canonical_url: "https://www.modelscope.ai/models/iamzwxyy/groot-mind2act"
md_url: "https://www.modelscope.ai/models/iamzwxyy/groot-mind2act.md"
repository: iamzwxyy/groot-mind2act
last_updated: 2026-11-10
license: other
model_type:
  - Gr00tN1d7
architectures:
  - Gr00tN1d7
parameters: 3.5B
tensor_type:
  - BF16
library_name:
  - safetensors
downloads: 5
stars: 0
tags:
  - robotics
  - vision-language-action
  - GR00T
  - Mind2Act
  - arx-x5
---

# groot-mind2act

> groot-mind2act - An open-source model by iamzwxyy on ModelScope. GR00T Mind2Act final non-EMA full-model inference weights, step 64000

iamzwxyy/groot-mind2act is a 3.5B-parameter machine learning model on ModelScope. licensed under other.

- **Repository**: iamzwxyy/groot-mind2act
- **License**: other
- **Parameters**: 3.5B
- **Tags**: robotics, vision-language-action, GR00T, Mind2Act, arx-x5
- **Downloads**: 5
- **Stars**: 0
- **Last updated**: 2026-11-10

Source: https://www.modelscope.ai/models/iamzwxyy/groot-mind2act

---

# GR00T-N1.7 · Mind2Act

GR00T-N1.7-3B fine-tuned on the Hugging Face published **Mind2Act 4500-episode** training set. This release contains the **final non-EMA full-model weights at optimizer step 64,000**, native processor/statistics, the matching GR00T runtime and a small inference loader.

Training used **16 A800 GPUs**, global batch **640**, per-GPU microbatch **20**, gradient accumulation **2**, approximately **5 equivalent dataset passes**. Language and vision backbones were frozen according to the upstream configuration. Both workers exited successfully. All 1,031 saved tensors are finite, and the final model files match the last periodic checkpoint byte for byte. This release contains no optimizer state, raw training episodes or videos.

## Input and output

- Three current RGB `uint8` HWC images: `cam_high`, `cam_left_wrist`, `cam_right_wrist`.
- Current `float32` state of shape `(14,)`: left arm joints × 6, left gripper, right arm joints × 6, right gripper.
- A task instruction string.
- Output: **16 × 14 absolute joint/gripper targets** in the same order. Joint angles are radians; grippers retain the dataset's controller units.

The learned actions are relative to the current state and use native percentile normalization, including grippers. The supplied loader delegates decoding to the saved native processor, which returns absolute targets; **do not add the current state a second time**. Training observations are 30 Hz. `config.json` retains the architecture's maximum horizon of 40; the saved `NEW_EMBODIMENT` processor defines the actual 16-step action chunk.

## Installation and download

Use Linux, Python 3.10 and a CUDA GPU. The verified environment used PyTorch 2.7.1/cu128, torchvision 0.22.1 and transformers 4.57.3. Exact inference package versions are in `inference_environment.json`.

```bash
pip install torch==2.7.1 torchvision==0.22.1 --index-url https://download.pytorch.org/whl/cu128
pip install -r requirements.txt
```

Flash Attention is optional: the bundled upstream code falls back to SDPA if it is unavailable. The GPU release check used flash-attn 2.7.4.post1. Use an official wheel matching Python 3.10, Torch 2.7 and your C++ ABI if installing it.

Download this repository with ModelScope or Git LFS. Run inference from its root, or import its `inference.py` by absolute path. The loader puts the bundled `runtime/` on the Python path; do not substitute an unrelated installed GR00T version.

The native backbone constructor still loads **Cosmos-Reason2-2B** before the complete fine-tuned checkpoint is strictly restored. It is a separate dependency, not another fine-tuned checkpoint. Use an existing local snapshot or obtain access to the upstream repository and download the pinned revision:

```python
from huggingface_hub import snapshot_download
cosmos_path = snapshot_download(
    "nvidia/Cosmos-Reason2-2B",
    revision="9ce19a195e423419c349abfc86fd07178b230561",
)
```

Cosmos requires its upstream access terms; this release does not provide credentials. Pass the resulting local directory explicitly to run without network access:

```python
from inference import load_policy

policy = load_policy(cosmos_path="/path/to/Cosmos-Reason2-2B", device="cuda:0")
actions = policy.predict_action(
    images={
        "cam_high": head_rgb,
        "cam_left_wrist": left_wrist_rgb,
        "cam_right_wrist": right_wrist_rgb,
    },
    state=current_state_14,
    instruction="Your task instruction",
)
assert actions.shape == (16, 14)
```

The loader checks missing, unexpected and mismatched model keys and overrides machine-specific Cosmos paths in memory. It does not rewrite checkpoint files. Model files in the original training directory remain intact.

## Validation scope

See `parameter_validation.json`, `training_summary.json` and `release_validation.json`. Release validation reloads the saved model and runs the native four-step action sampler on real observations at frames 0 and 30, checking action shape, finiteness, modality/order, relative-to-absolute decoding and unchanged configs. This is a functional inference check, **not a held-out task-success evaluation**; no robot rollout or benchmark success rate is claimed.

Earlier preparation also passed isolated two-GPU training, no-grad validation and checkpoint resume. This publication does not modify training code or rerun formal training.

## Sources and license

- Base: [nvidia/GR00T-N1.7-3B](https://huggingface.co/nvidia/GR00T-N1.7-3B), revision `2fc962b973bccdd5d8ce4f67cc63b264d6886495`.
- Backbone: [nvidia/Cosmos-Reason2-2B](https://huggingface.co/nvidia/Cosmos-Reason2-2B), revision `9ce19a195e423419c349abfc86fd07178b230561`.
- Training/runtime: [XPolicyLab](https://github.com/XPolicyLab/XPolicyLab), revision `0075cc96a1bf24c5d66a4a3a5e29eae6d739fef3`, with the Mind2Act preparation fixes retained in the bundled code.
- Dataset: [Mind2act/mind2act-train](https://huggingface.co/datasets/Mind2act/mind2act-train), revision `2d54f52c77d4cda85e340bd14a034cd0c377d196`; 8,191,528 frames, published original CP01 version.

**Weights: NVIDIA License, research/evaluation use only, as stated in the pinned base snapshot's `LICENSE` §3.3.** Its model card contains broader commercial-use wording; this release preserves the actual LICENSE without weakening that restriction. Read the included `LICENSE` for the complete terms. GR00T code is Apache-2.0 (`LICENSE-code-Apache-2.0.txt`); upstream attributions are preserved in `ATTRIBUTIONS-GR00T.md`. Cosmos components are licensed by NVIDIA Corporation under the NVIDIA Open Model License; its original model card and terms are preserved under `upstream/`. Separate component licenses continue to apply.

This is a Mind2Act fine-tuned derivative, not an official NVIDIA release. Full provenance is recorded in `SOURCE_INFO.json`; file integrity hashes are in `SHA256SUMS`.