Background
My VisionX internship involved vision-guided cloth folding, connecting multimodal observations, policy actions and a simulation environment. This page describes my own contributions to that workflow.
Goal
Build a runnable, inspectable local vision-to-action validation workflow and provide reusable simulation components for subsequent folding-policy integration.
My contributions
- Deployed X-VLA on Linux, addressing dependencies, multimodal observations, policy outputs and integration constraints.
- Implemented client/server inference and motion-trajectory visualization.
- Trained and evaluated on official demonstrations, comparing predictions before and after training to inspect the perception-to-action chain.
- Fixed cloth stiffness and configuration issues in company simulation code and packaged the cloth component for handover.
Approach
- Replayed dataset trajectories before introducing predictions to distinguish data, action-mapping, cloth-physics and simulator issues.
- Used MuJoCo trajectory visualization and a basic Isaac Sim cloth scene for staged checks.
- The workflow involved cross-embodiment soft-prompt adaptation; the precise implementation scope and experimental configuration remain to be documented.
Results
- Delivered local inference, trajectory visualization and a staged validation workflow.
- Handed over a reusable cloth component for subsequent folding-policy integration.
The resume provides neither folding success rates nor real-robot validation results. This page describes workflow and component deliverables only.