Real policies.
Measurable dexterity.

Measure the policy and a human operator on the same embodiment, objects, task mix, fixtures, observation setup, and rules. For pick and place, the human teleoperates the same arm + hand system. Hand-only tasks use the same mounted hand and fixture.

For each run, throughput is successful task completions divided by total elapsed run time. Start timing at the first trial cue and stop when the last planned trial ends. Failed attempts, timeouts, recovery, and between-trial resets count toward elapsed time. Predeclare any excluded downtime and apply the same rules to both runs; do not time successful trials alone.

HRT (%) = 100 × (policy successes / policy elapsed time) / (human successes / human elapsed time)

100% matches the human’s completion rate; 50% is half; 200% is twice. This is a direct normalization to the human teleoperation baseline. Report success rate alongside it so readers can distinguish reliability from speed. A reference with zero successful completions cannot define the ratio.

Record the operator, experience, trial counts, timing scope, and exact hardware. Repeat sessions to estimate uncertainty; a single aggregate run cannot establish a throughput confidence interval. Keep object weights fixed for multi-object comparisons. The current submission format does not include full-run elapsed time or a matched human reference, so HRT requires additional run records.

Each release must identify its task, protocol revision, scene set, scoring rules, time cap, and included bodies. Freeze the policy checkpoint and disclose every embodiment adapter. If a body requires different fine-tuning, label that as an adapted-policy comparison; it is not a zero-shot transfer result.

The proposed publication floor is 40 valid trials per body × policy × task cell, collected through repeated complete protocol rounds. Publish all exclusions and their reasons. This is an eligibility floor, not a claim that 40 trials establishes statistical significance. Single 20- or 30-trial runs can appear in community results.

Cross-body summaries require at least three eligible bodies and use the intersection of eligible bodies across all policy columns in the selected controller group, so the displayed summaries share the same coverage. Report population standard deviation, mean success, the included body count, and the exact release manifest. HRT additionally requires complete-run timing and a matched human teleoperation reference for every included cell.

Only a reviewed RC release supplies the official matrix. Anyone can reproduce the calculations from its published records; submitting a run does not confer RC verification.

A dash means unmeasured or unavailable, never zero. N/A means the task requires an ability absent from that proposed configuration, such as three independently assigned fingers. An insufficient cell has too few valid trials for an RC summary.

Success-rate intervals use the 95% Wilson interval for independent binary trials. Correlated trials require episode-aware analysis before RC publication. Hardware-invalid trials must be justified and counted separately; policy errors, unsuccessful attempts, and timeouts remain in the denominator. Community completion times cover only trials with a reported time and are not a throughput score or ranking tie-breaker.