Data, Hardware & Evaluation
The embodied data bottleneck and the machines themselves, with the measurement crisis behind every benchmark claim.
This domain overview lists 7 articles, ordered as where embodied training data comes from, then the rigs that produce it, then why two reported benchmark numbers are usually not comparable at all.
Robot-hours versus LLM tokens: the log-log reality of embodied data and teleop-farm economics.
Open X-Embodiment, DROID, BridgeData V2, AgiBot World, RoboMIND: five datasets compared.
Arms, humanoids, hands, sensors, and compute: a buyer's guide from SO-101 to Jetson Thor.
ALOHA, GELLO, UMI, and VR teleop: cost, data quality, throughput, and the embodiment gap.
Why N-of-10 trials and unreported variance mislead: 95% per-step success is unusable at 30 steps.
The installed base robot learning is trying to enter, and the jam-rate arithmetic that decides whether a 99 percent cell ships.
Data capture, schemas, training, simulation, evaluation, serving and robot integration as one reproducible system rather than a model checkpoint.