NTTドコモ builds robot ML pipeline in 3 days with AI-DLC

The backdrop here is a familiar pain in applied machine learning: research papers rarely transfer cleanly into working pipelines. Small differences in assumptions and implicit conditions surface only after implementation, and experiments that fail are expensive. NTT Docomo's team, working on physical AI (AI that acts in the real world), had run into exactly this — its imitation-learning model reached a 10% success rate on 10 demonstrations, rose to 45% with data amplification, then plateaued.

The team's choice was to treat the pipeline not as a coding exercise but as a product to be structured with AWS's AI-DLC framework, which places AI at the center of planning while keeping humans on decisions. In the Inception phase, the team fed papers and domain material into the Kiro tool and worked in a mob style around it. The AI review flagged issues like a checkpoint-format mismatch, a 12D-versus-28D input-dimension conflict, and inconsistent action scaling before any code was written — exactly the class of error that tends to appear later in ML work as "the code runs but the output is wrong."

What this shows so far is a first installment: the Inception phase, not the final training results. The genuine test will be the Construction phase, where two teams ran in parallel — one reportedly abandoning its original scope and pivoting within the AI-DLC frame. Whether the structured-document approach actually improves ML outcomes, rather than just improving coordination, is what the second part seems positioned to answer.