Overview
When ChatGPT emerged in 2023, I immediately connected its chat context with robot APIs to build an MVP for natural human-robot interaction in inspection robots. During experimentation, an emergent pattern appeared — the robot began "creating behaviors" on its own, subtly demonstrating the possibility of robotic self-evolution. The system was built on DAMO Academy's "XiaoJinNiu" and "WALL-E" platforms, both power grid autonomous inspection robots — one indoor, one outdoor.
My Role
AI Natural Interaction Technical Lead. Responsible for designing the RoboChat framework, including LLM-based robot brain design, response processing, executable Python code extraction, environment creation, API invocation, and event dispatching.
Key Decisions
- Leveraged the LLM's "understanding capability" to generate and compose behaviors, rather than relying on traditional keyword-based detection and recognition.
- Built a pipeline: LLM content generation → executable code extraction (early tool use) → API invocation → action feedback loop.
- Deployed in real-world robot application scenarios and integrated into operational workflows.