Perception servers

TiPToP asks two servers on every rollout: M2T2 for grasps and FoundationStereo for depth. tandem init builds both, with torch compiled for your GPU and their model weights. You don't need to start them yourself: tandem collect starts any server that isn't running before the session warms up, and stops the ones it started when the session ends.

tandem servers status   # installed? answering?
tandem servers start    # start them ahead of time; they keep running until `tandem servers stop`

If the perception servers are on another machine, point TANDEM at it:

tandem rig set planners.tiptop.perception.m2t2.url http://HOST:8123
tandem rig set planners.tiptop.perception.foundation_stereo.url http://HOST:1234