Relari

Relari's core claim is that a hand can look still in video while its muscles continuously regulate force, contact, and stability — and that the missing signal is exactly what makes human manipulation so hard to copy. Its answer is surface EMG: a single lightweight armband that adds interaction information to natural human demonstrations without instrumented objects, gloves, or robot hardware. The site names four specific signals it extracts: grip force distribution across fingers, stiffness from muscle co-contraction, anticipation of force before contact, and rapid reflex responses to slips and mistakes. Those are the details that matter, because they are the parts of manipulation that vision-only datasets tend to miss. The scalability argument is the interesting one. Relari frames the field as a tradeoff: teleoperation is richest but lowest scale, video plus tactile glove is very rich but low scale, and video alone scales highest but carries the weakest signal. It places EMG plus video between those poles and frames the goal as moving force supervision from a robot, operator, and lab to people across more tasks and environments. If that works, it is a genuinely larger data pipeline than glove-based collection. The honest weaknesses start with maturity. There is no published pricing, no pricing page content, no documented integrations, and no public product or API described in the scraped material. What the site does show is a research narrative and an open hiring call for people with deep expertise across robot learning, human biomechanics, and hardware. That means any evaluation today is an evaluation of a research team and thesis, not of deployable software. There is also real technical risk in the approach itself. EMG signals vary across people, sensor placement, and skin conditions, and the site does not publish accuracy figures, cross-subject generalization results, or benchmarks against teleoperation. The transfer step — decoding biomechanical structure and re-targeting it across robot embodiments — is asserted rather than demonstrated in the scraped content. Until there are published results, treat the embodiment-transfer claim as a hypothesis. Where it fits: academic and industrial robotics labs working on manipulation foundation models, biomechanics groups with EMG instrumentation experience, and investors tracking physical AI data pipelines. Where it does not: anyone who needs a production manipulation API this quarter, teams without EMG or biomechanics capability, and buyers who need published pricing or integration documentation to run a procurement cycle.