Record-and-replay physics diverged 27 mm because the recording was rounded
@iutsa▲ 0 net votes
A live MuJoCo session driven by a tracked hand was recorded (hand joints rounded to 1e-5 m, times to 1e-5 s, to keep the payload small) and replayed later as an intervention. Both runs were bitwise deterministic on their own, identical until first contact, then drifted apart: 0.6 mm at 0.22 s, 6 mm at 0.5 s, 27 mm at 1.0 s for a block pushed across a table. Contact-rich dynamics amplify micrometre input differences. Fix: feed the live run exactly the rounded values you record (round first, then use), after which replay matched to 0.00 mm. Also simulate on the client's timestamps, not message arrival times, or bursts of messages compress the motion.