---
title: TRACE
canonical_url: "https://www.modelscope.cn/models/kangzejian/TRACE"
md_url: "https://www.modelscope.cn/models/kangzejian/TRACE.md"
repository: kangzejian/TRACE
chinese_name: TRACE
last_updated: 2026-10-21
license: apache-2.0
pipeline_tag: robotics
tasks:
  - robotics
parameters: 132.1M
tensor_type:
  - F32
library_name:
  - safetensors
language:
  - en
downloads: 6
stars: 0
tags:
  - text-to-motion
  - unitree-g1
  - diffusion
---

# TRACE

> TRACE - kangzejian 在 ModelScope 开源的模型。Typed Reference-Aware Conversational Editing for Humanoid Motion Diffusion. Text-to-motion generation and reference-aware editing for Unitree G1.

kangzejian/TRACE 是 ModelScope 魔搭社区上的 132.1M 参数robotics模型，采用 apache-2.0 许可。

- **Repository**: kangzejian/TRACE
- **License**: apache-2.0
- **Tasks**: robotics
- **Parameters**: 132.1M
- **Tags**: text-to-motion, unitree-g1, diffusion
- **Downloads**: 6
- **Stars**: 0
- **Last updated**: 2026-10-21

Source: https://www.modelscope.cn/models/kangzejian/TRACE

---

# TRACE: G1 inference checkpoints

TRACE checkpoints are released under Apache-2.0. Upstream components retain their own licenses.

**TRACE: Typed Reference-Aware Conversational Editing for Humanoid Motion Diffusion**

[arXiv paper](https://arxiv.org/abs/2610.21515) | [PDF](https://arxiv.org/pdf/2610.21515) | [Code, method figure, and MP4 examples](https://github.com/Multimodal-Perception-Action/TRACE) | [Inference resources](https://modelscope.cn/datasets/kangzejian/TRACE-Inference-Assets)

Contents: original sparse-event generator EMA, idle-boundary generator EMA, a separate source-relative hand-height editor, and two separate PEFT LoRA adapters for generation planning and edit parsing. Both adapters use Qwen3-4B-Instruct-2507. The base language model and Qwen3-Embedding-4B are obtained from their official repositories, not duplicated here. The model companion code and inference-resource repository are required; no hosted ModelScope `pipeline()` integration is claimed.

The `edit.py` entry point edits 20 FPS G1 source clips of at most 104 frames. It supports one relative left/right hand-height request per call, up to 10 cm, with an explicit time interval. Protected coordinates are copied from the source throughout sampling; no editing IK or post-smoothing is used. See the companion repository's reference-and-editing guide.

Input: English text. Output: 20 FPS Unitree G1 kinematic motion. The source implementation uses 40D gravity-gauged states, raw-text embeddings and numeric sparse-event conditioning. Default inference includes endpoint conditioning, reference projection, smoothing and geometric heel/toe refinement. Physical contact labels in this postprocessing branch are rule-derived, not an independently trained four-region classifier.

The models generate flat-ground kinematic references. Compound-action sequencing and foot contact remain challenging. Robot execution requires a separately validated tracking controller.

Training data includes Motion Data by [Bones Studio](https://bones.studio/). Use of the underlying dataset is subject to the BONES Motion Capture Dataset License Agreement. No raw or retargeted dataset motion is included.

Artifact hashes, file paths, and installation commands are supplied in the companion code's `release_manifest.json` and README.