stream-to-act

KO | EN

gitlite — search

#python #java #python3 #arduino #golang #machine-learning #rust #html #flask #javascript #seismology #nodejs

★ 10 Open GitHub ↗

ROS 2 Native Token Streaming for Smooth, Real-Time Control Loops (ICRA 2026)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# 🚀 Stream-to-Act  
ROS 2 Native Token Streaming for Smooth, Real-Time Control Loops

Dahyun Kim, Yunseong Jeon, Hongkyun Park, Jong-Chan Kim  
2026 IEEE International Conference on Robotics and Automation (ICRA)  
Accepted

---

## 📌 Overview

Stream-to-Act is a ROS 2-native execution framework that enables
token-level streaming from Large Language Models (LLMs) directly
into real-time robotic control loops.

Conventional LLM-driven robotic pipelines wait for full-sequence
generation before action execution, introducing avoidable latency.
Stream-to-Act instead executes actions incrementally as tokens
arrive, reducing control delay and improving responsiveness.

---

## 🧠 Key Idea

Instead of:

LLM → Full Response → Parse → Execute

Stream-to-Act performs:

LLM → Token Stream → Incremental Execution → Continuous Control

This streaming architecture enables:

- Reduced end-to-end latency
- Improved temporal smoothness
- Lower actuation jitter
- Compatibility with ROS 2 executor model

---

## 🏗 Architecture

(Insert system diagram here)

The framework consists of:

- A ROS 2 streaming executor
- A token-level scheduling module
- An incremental action dispatcher
- A closed-loop control interface

---

## 📂 Repository Structure

```
stream-to-act/
├── stream_executor/     # ROS 2 streaming executor
├── token_scheduler/     # Token-level scheduling
├── demo_waypoint/       # Waypoint driving demo
└── evaluation/          # Latency & jitter analysis
```

---

## ⚙️ Installation

Core framework setup:

- ROS 2 (Humble or later)
- Python 3.10+
- CUDA-enabled GPU (recommended)

Simulation environment (CARLA) is provided in a separate repository:

👉 https://github.com/홍균깃헙

---

## 🔁 Reproducibility

Tested on:

- Ubuntu 22.04
- ROS 2 Humble
- NVIDIA RTX series GPU
- Local LLM backend with streaming enabled

---

## 🔗 Acknowledgment

This project builds upon the open-source project  
[llama_ros](https://github.com/mgonzs13/llama_ros) (MIT License).

The original repository provides a ROS 2 integration for LLaMA-based
models. We significantly modified the execution pipeline to enable
token-level streaming and incremental action dispatching for
real-time control loops.

---

## 📖 Citation

If you find this work useful, please cite:

```bibtex
@inproceedings{kim2026streamtoact,
  title     = {Stream-to-Act: ROS 2 Native Token Streaming for Smooth, Real-Time Control Loops},
  author    = {Kim, Dahyun and Jeon, Yunseong and Park, Hongkyun and Kim, Jong-Chan},
  booktitle = {Proceedings of the 2026 IEEE International Conference on Robotics and Automation (ICRA)},
  year      = {2026},
  note      = {Accepted.}
}
```

stream-to-act

Explore Similar Repositories

Was this content helpful?

Recent Feedback

Download README