General-purpose robots need to handle a wide range of objects and tasks, not just a single setup. The new arXiv preprint introduces HandsOn, a composable simulation platform and benchmark built specifically for cross-embodiment dexterous manipulation. The authors argue that reliable behavior must hold across embodiments, objects, and task demands, and HandsOn is designed to test that.
The abstract describes HandsOn as large-scale and composable, but the full details are not available in the announcement. What is clear is the focus on generalization: rather than evaluating a single robot arm on a single task, the benchmark is meant to support research that transfers skills across different hardware and object sets.
The paper is a single source, so there are no conflicting findings to compare. The contribution appears to be a shared evaluation tool for the dexterous manipulation community, though the specifics of the simulation environment and benchmark tasks remain to be detailed in the full paper.