PiPER grip clock · v0.3 · replay study
The clock between commands.
A small learned clock decides when to wait and when to move. Compare it with a clock trained without live grip measurements, fixed timing and an authored teacher. Every scene uses the same original arm and jaw references. These are recorded simulation results; physical robots and learning from human videos remain future work.
The frame timeline follows fixed physical simulation time. Clock progress and wait debt are separate recorded outputs. “Clock output” is the model's advance probability, not a safety confidence. Terminal frames have no invented next action. Failed attempts remain included, including outcomes with no saved native frames. Playback uses saved states and does not run new physics or inference. The complete source evidence is linked below.
All seven arm and jaw references are frozen. The clock without grip measurements keeps the same starting scene, time features and movement limits. Every simulation uses the same 20 ms control interval. The selected variant chooses whether virtual progress holds or advances. No camera or physical robot has been tested here.
All scheduled outcomes
Loading complete outcome index…
Inspect all 512 attempts and their evidence
Evidence and limitations
The outcomes cover 128 cases across four variants (clock, blind clock, frozen fixed, and authored teacher), with nominal and post-grip conditions. Replay selection does not depend on success. Pinned report, native arrays, receipts, score and start records are linked for every outcome. The offline exporter pins dataset links to one exact revision. The publication receipt records upload and anonymous readback verification separately.