Simbo

• 

  Main Page

• 

  Programming agent

• 

  World design agent

• 

  End product by Simbo

Inspiration

As someone who actively participates in robotics competitions, I know simulation is crucial for testing algorithms safely before deploying them on real robots. However, Gazebo and ROS are not beginner-friendly, and robotics engineers often spend most of their time troubleshooting. This inspired me to create a multi-agent system to help robotics engineers, learners, and students quickly generate simulations and control programs.

What it does

• Accepts one prompt, searches for/open-source map models, and imports the best match.
• Analyzes your ROS/ROS2 workspace structure and existing code.
• Generates control programs (velocity/position/joint/teleop), launch files, and workspace edits.
• Provides an interactive Streamlit UI and programmatic API.

How we built it

• Orchestration with LangGraph + LangChain.
• Workspace tools to read, analyze, and edit ROS workspaces.
• OpenAI models for code generation and iterative error-fix reflection.
• Streamlit frontend for easy interaction.

Challenges we ran into

Fine-tuning the agent was really challenging, especially since we needed to consider almost all the factors that could go wrong during a ROS workspace setup and adjust the agent for them to aoid.

Accomplishments that we're proud of

Completed programming the agent and map agent to operate completely autonomously, helping users generate the map files and control programs they need using just one prompt. The overall workflow of the agents allows them to keep reflecting, fix system errors, and output a well-constructed workspace to the user.

What we learned

Fine-tuning LLMs, multi-agent system design, LangGraph design, prompt engineering, and Streamlit.

What's next for Simbo

Include the remaining two agents with a manager agent to complete the overall workflow, allowing an entire workspace to be set up with one prompt.

Built With

• langchain
• langgraph
• python