Language-guided humanoid manipulation
A simulated Unitree G1 that turns varied language instructions into grounded, validated manipulation actions inside MuJoCo.
- Integrated a locally deployed Qwen2.5-VL-7B perception pipeline with OpenCV refinement.
- Built structured planning, inverse-kinematics motion, contact-aware grasping, recovery, and visual verification.
- Created reproducible scenario evaluation, JSONL evidence, and 33 automated tests.
- InstructionA natural-language task enters the system.
- PerceptionQwen2.5-VL and OpenCV ground the object and scene.
- PlannerThe grounded task becomes a manipulation plan.
- G1 controlThe humanoid reaches, grasps, transports, and releases.
- VerificationThe result is inspected and recorded as evidence.