Thea & Unitree G1
Deploying Unitree G1 locomotion and manipulation within a closed-loop embodied-agent harness for long-horizon tasks and failure recovery.
Thea brings the harness paradigm from coding agents into the physical world. It exposes robot capabilities as callable tools and closes the loop with structured world state, action termination, outcome verification, and failure diagnosis.
I led the deployment, integration, and debugging of Unitree G1 locomotion, grasping, and pick-and-place capabilities, including robot information access and system configuration for multi-step tasks in real environments.
I also contributed to:
- a closed-loop tool-calling framework that unifies language models, robot policies, and physical tools for dynamic long-horizon task composition and failure recovery;
- a persistent scene graph that provides a queryable, structured representation of the world; and
- independent evaluators that classify actions as successful, failed, or in progress and feed the result back into subsequent decisions.
Project page · arXiv preprint · PDF
