Projects
Ongoing
Cost-Effective Leader Arms for Teleoperation
Benchmarking the most cost-effective and performant leader arms per dollar for robot teleoperation. Comparing affordable builds such as GELLO and Koch-style arms to identify the best options for low-budget research labs looking to collect high-quality demonstration data. The evaluation also factors in ergonomic design to prevent overuse injuries for operators during extended data collection sessions.
Ongoing
Auto-Labeling for Robotic Data Collection
Building software that automatically generates labels for robotic data collection, covering both visual annotations such as object detection and segmentation, and behavioral labels including task segmentation and success or failure classification. The labeled data feeds into a reinforcement learning pipeline to train reward models and fine-tune vision language action models with diverse language instructions and proprioceptive state.