A robot policy is only as good as the data underneath it. Almost nobody checks the data. We do — episode by episode, tracing each defect from the broken record back to the broken behavior.
Trace Machines is independent robotics research. Right now the work is forensic: auditing the datasets that robot foundation models are trained on, and tracing the defects hidden inside them. It's where we're starting, not the edge of what this becomes.
Go through episodes directly instead of trusting the summary stats. Assume the dataset is lying until checked.
Find where a flaw originates, which episodes it touches, and how it would corrupt training downstream.
Neutral and buyer-side, with reproducible evidence. No incentive to flatter the data.
A hands-on, reproducible audit of a widely-used robot learning dataset. Findings include blank-but-recoverable language annotations, truncated annotation coverage, and unmarked failure episodes.
github.com/goelankur1996/droid-auditNotes on robotics and machine learning — thinking out loud as the work goes.
lessclueless.me