From July to September 2026 I filled in on a RoboService team working to deliver box packing and manipulation skills on an Agibot A2 humanoid robot for Siemens. I advised on formalizing the team's model development pipeline so they could stop guessing which data works and start shipping working robot models.
The video above (shown at 2x speed) shows the working task end to end: the robot separates and picks packaging material, then performs the external pickup of the packed box. The first segment runs through separator handling; the second shows the pickup executed with the external setup.
The team had the hardware and the ambition but was spending most of its time discovering what data would train. I brought in a structured pipeline: what to collect, how to process it, and which training techniques to apply with different models. Concretely:
The team went from trial-and-error data gathering to a repeatable pipeline and delivered working robot models to Siemens on schedule. The video is a demo of the final behavior running on the physical Agibot A2.