DexArena benchmark tests two-handed robots on long chores
Researchers built a test to see if robots can finish whole chores.
A new benchmark tests robots on whole chores, not just single moves.
It's called DexArena, and it was built to evaluate learning and generalization in long-horizon bimanual dexterous manipulation. In plain terms: robots using both hands to chain several skills into one long job. That is difficult to evaluate systematically.
DexArena includes 10 long-horizon tasks and 30 stage-level tasks, set in household and industrial scenarios. Each long task is split into explicitly defined stages, and each stage has a physically grounded success condition. That means researchers can evaluate both intermediate progress and end-to-end task completion.
The researchers say existing manipulation benchmarks mostly stress isolated skills or support only a limited range of end-effectors. DexArena targets bimanual dexterous manipulation, where the abilities in play include skill composition, bimanual coordination, generalizable visual understanding, and robust transitions between skills.
The benchmark comes with a unified interface and verified human-teleoperated demonstrations, so different policy learning algorithms can be tested. It also provides standardized simulation configurations, scene randomization, trajectory replay, and task-level evaluation.
In their baseline experiments, the team identified substantial challenges in precise manipulation, generalization to scene variation, and skill composition.
The paper was submitted to arXiv on 2 November 2026 and lists a project page.
What chore would you most want a two-handed robot to finish start to end?
Archive video: Cho E, Chen R, Merhi L, Xiao Z, Pousett B, Menon C, CC BY 4.0 (Wikimedia Commons); Marcos Gonzalez, CC BY 4.0 (Wikimedia Commons); NASA
Video: sipair-elastic.github.io/DexArena-website