Towards Reducing the Barrier to Data Collection
Towards Reducing the Barrier to Data Collection
We enhanced the data collection pipeline for LeRobot by addressing two key limitations: lack of depth data and absence of a real-time simulation model. Our work improves both the robustness of Vision-Language Agents (VLAs) and the speed of dataset generation.
Key Contributions:
Depth Sensor Integration:
We incorporated an Intel RealSense depth camera into the LeRobot system, enabling asynchronous capture of RGB-D data from three perspectives: the gripper camera and two external views. This expands the existing setup—which only supported RGB and text—to include depth, a modality shown in prior research to significantly improve VLA performance.Digital Twin Implementation:
We developed a digital twin that mirrors the physical robot's actions in real time during teleoperation. This provides a simulated view for enhanced monitoring, debugging, and potential offline training in future iterations.
Our system boosts data richness and lays groundwork for more scalable and reliable VLA training and evaluation.