Research
What we are measuring, and what it has told us so far.
An ongoing pilot on a few-hundred-dollar SO-101 arm, asking one question: can demonstrations recorded on a handheld gripper replace part of the robot teleoperation normally needed to train a manipulation policy? These are the write-ups, in order. Each reports its intervals and names what it does not show.
-
September 2026 · latest
Can handheld video replace teleoperation?
The pipeline end to end on a physical robot: handheld GoPro demonstrations converted into joint trajectories, executed on the SO-101, recorded as robot-native episodes, and used to train an ACT policy. Sixteen teleoperated demonstrations reach 16%. Holding those fixed and adding sixteen filmed by hand reaches 67% — against 72% for doubling teleoperation outright.
-
August 2026
What 81 robot mappings taught us about better robot data.
The screening work this pilot is built on. One measured human trajectory mapped onto a five-joint arm 81 different ways, with every frame inspected, to find which frames fit the embodiment and which needed correction before the motion was worth putting into a learning loop.
The second piece is not superseded by the first. One asks whether a mapping is good; the other asks whether good mappings are worth anything to a policy. They are separate questions and we needed both answers.