{GSoC 2026} Project #10: Agentic Dora: Agent-Driven Robot Control

[fayezzouari]

Hi,

I am interested in the "Agentic Dora" GSoC project idea and wanted to share what I've built so far and get your feedback before finalizing my proposal.

I implemented a proof-of-concept that bridges a natural language agent with Dora's dataflow system to control a simulated robot. Here's what it includes:

• A three-node Dora dataflow: cli_interface, agent_bridge, and simulation
• A kinematic robot simulation with optional PyBullet GUI (PYBULLET_GUI=1)
• A mock agent (no API key required) for offline testing, with keyword-based command parsing
• An LLM-powered agent using smolagents with three supported backends:
  • Ollama (default, fully offline): runs qwen3:1.7b locally via Ollama's OpenAI-compatible endpoint
  • Groq: llama-3.3-70b-versatile via GROQ_API_KEY
• Real-time command dispatch via a shared queue so each tool call takes effect immediately during agent execution
• Closed-loop turn control using heading feedback from robot_state
• A robot CLI for sending natural language commands (./robot "move in the shape of a square")
• 20 passing unit tests

The agent can handle commands like "move forward for 5 seconds", "turn left", "move in the shape of a square" (LLM plans and executes the full sequence), and "status".

Before writing the full proposal I wanted to confirm: is this the right direction? Should the simulation remain PyBullet-based, or is there a preferred robot/simulator? Is the agent framework flexible? And are there specific Dora features you'd like the project to demonstrate? Is there a proposal template that I should follow?

You can find, attached to this discussion ,a video demo and the link towards the dedicated github repository!

Github Repository: https://github.com/fayezzouari/dora-agent
Demo Video: demo-dora-agent.webm

Looking forward to your feedback!
Thanks

[bobdingAI]

The bridge between a natural language agent and Dora's dataflow is right direction. For simulation we don't have a hard preference. One possible Dora's value is showing how nodes can be added or swapped without rewriting the pipeline. There isn't a strict template, but a good proposal should cover: the problem and motivation, your proposed architecture and timeline etc.

[fayezzouari]

Thank you for the feedback, I sent my proposal to the org email address and looking forward to your review!