2026 Spring Robotics Colloquium: Parker Owan (Velodex Robotics)

In this talk

1. 2:28

   Lessons from Amazon and Cruise

   At Amazon, Owan helped create Vulcan Stow, which used a purpose-built gripper, touch sensing, and a set of manipulation routines to place inventory in crowded pods. At Cruise, he worked on choosing vehicle trajectories using machine learning and learned to prioritize rare, consequential driving events. He says manipulation opportunities often involve costly, varied work, and that deployment, exception handling, and careful evaluation matter as much as the core task.

2. 6:37

   Why food production is a robotics opportunity

   Owan chose food manufacturing because it combines a large market with hard technical problems that can be tackled incrementally. He cites customer research indicating 80% annual labor turnover, 20% product waste, and business margins of 5% or less. Large automated lines serve high-volume products, but many lower-volume tasks remain manual because dedicated machines are costly and underused.

3. 14:46

   Designing robots for kitchens

   Velodex’s approach combines hardware, reactive control, policy learning, and deployment. Owan describes experiments with grippers for tasks such as handling delicate fruit, arranging cups, and sealing lids, as well as portable single-arm and dual-arm platforms designed to reach common work surfaces. Food settings also demand cleanable designs, food-safe materials, protected wiring, and sensing that captures contact and force as well as visual information.

4. 27:23

   Learning and deploying manipulation policies

   Owan argues that force-aware control helps robots manage contact, while learned policies are most useful for the final few centimeters of a task. In a cup-placement experiment, success rose from 10% with 50 demonstrations to about 65% as the team changed its model and training approach; he highlights modality dropout and depth-specific features as useful improvements. Replay, offline analysis, closed-loop tests, and hardware trials helped the team distinguish model errors from timing and compute problems.

5. 45:44

   A production path and next steps

   Velodex plans to refine its system through real production work, using customer workflows to gather useful data and expand the range of tasks. Owan describes a possible 30-to-40-robot co-development facility where food producers and technicians would work alongside robots during a supervised ramp-up. Longer-term plans include learning from deployment, collecting 50,000 hours of combined demonstration and reinforcement-learning data, and building a new robot platform.