From Demos to Deployment for Physical AI | Edward Liu (RealMan)

In this talk

1. 0:26

   Deployment means a complete workday

   Liu contrasts a robot completing a task once with performing it reliably across four shifts, despite changes in objects or lighting. A deployed system must handle failures, recover, and produce results customers accept. With larger fleets, teams also need shared records of robot status, software, tasks, and responsible operators.

2. 2:29

   Use human support as autonomy grows

   Liu describes four connected needs: physical execution, fleet operations, human support for difficult cases, and learning from real work. RealMan’s Global Labor Network connects remote operators to robots, allowing useful work to begin before full autonomy. As robots improve, people can shift from direct control to exception handling and supervising multiple robots.

3. 4:08

   Build hardware for repeatable production

   RealMan’s three layers are deployable hardware and manufacturing, robot operations, and real-world data and learning. Liu says its joint modules have passed up to 50,000 hours MTBF, and the company has capacity for 10,000 complete robot systems and up to one million joint modules annually. He presents the 7.2-kilogram RM65 arm, with its controller integrated, as an example of reducing the weight and integration burden of a complete robot.

4. 8:50

   Connect operations and qualified data

   The operations platform links each robot’s task, operator, software configuration, and outcome, while distinguishing device health from whether a task is blocked. Liu says raw robot records become useful training data only when camera images, joint states, actions, and outcomes are synchronized and quality checked. Failures and human interventions can guide diagnosis, targeted training, evaluation, and monitored releases.

5. 14:04

   Start with measurable customer tasks

   Liu describes deployments, pilots, supervised operation, and teleoperation across different markets, without claiming every application is fully autonomous. He proposes starting with one useful task and clear acceptance criteria, with customers, integrators, and model teams contributing alongside RealMan. Examples include factory machine tending and material handling, warehouse picking and sorting, and facility inspection.