Sinewgrid Robotics trains policies that move real hardware.

A language model can learn from a large share of the written web. A robot has no such source. Demonstrations recorded by a human operator are exact but slow and costly to collect, human video is broad but carries no record of the actions taken, and simulation is plentiful but only an approximation of real physics. Sinewgrid Robotics treats these three sources as a pyramid and uses each for what it does well. Pretraining draws on video and simulation at scale to give the policy a general sense of objects and motion. Imitation on teleoperated episodes then teaches precise actions, and reinforcement learning in post-training improves the policy on tasks that can be measured. The result is a vision-language-action policy for manipulation and whole-body control.

What matters at the end is what happens on real hardware. Each task family has a fixed held-out set, a checkpoint is scored on real robots, and the gap between simulation and the real result is reported next to it, so a good simulated score is never mistaken for a real one. The compute is uneven: simulation rollouts can run on interruptible capacity, training needs a pool with fast interconnect, and evaluation runs on its own pool so it never competes with training. We work in open formats, with logs in MCAP, scenes in OpenUSD and demonstrations in a LeRobot-style dataset layout, so a partner can bring data in and take it out. Every policy has defined stop conditions and is tested on held-out tasks before it moves hardware outside the lab. The first body we target is chosen with the first partner.