Event

LocoManiLab releases a unified benchmark for humanoid mobile manipulation

1 reports · 1 sources

Dewei Wang and other authors released LocoManiLab, a unified benchmarking platform for humanoid mobile manipulation covering data collection and generation, policy training, evaluation, and real-robot deployment. Its task set contains 234 atomic and composite tasks across five categories of household scenarios, accompanied by 51.5 hours of simulated teleoperation data. The platform's evaluation covers the performance of 12 policies under a shared controller across tasks, scene splits, and perturbation levels, 2 policies across 5 controllers, and studies of demonstration scale and data-source mixing. Three of the simulated mobile manipulation tasks closely match real-world environment settings, allowing three policies to be evaluated in both the simulation and real domains.