CAMEL: Communicative Agents for Mind Exploration of LLM Society

CAMEL is a multi-agent (Agent Society) research and application framework emphasizing Role-Playing, tool integration, and task automation.

My Contributions

Based on the CAMEL framework, I conducted embodied Agent task implementation and systematic evaluation:

Designed Agent Tools and Prompt schemes for embodied tasks, supporting Zero-shot (text-only task description) and One-shot (single demonstration: video + text, Demonstration-conditioned / In-context imitation) task execution settings.
On the RoboTwin Benchmark, the constructed Agent system demonstrated significantly superior task completion ability compared to existing representative VLA models under Zero-shot conditions on unseen tasks.
Under One-shot demonstration conditions, the system further improved success rates, validating the effectiveness of Agent-centric methods in embodied task generalization and rapid adaptation.