Matt Strong

Matt Strong

CS PhD at Stanford University

← Paper Digest

DexArena: A Unified Benchmark for Long-Horizon Dexterous Manipulation

Haonan Ping, Jianshu Hu, Yimin Wang, Yi Zhu, Pai Liang, Qiyan Jin, Qirui Bao, Ying Qian, Xinyang Ding, Hongfei Qiao, Tianle Xue, Yucheng Bai, Soo Wei Jie, Shuguang Li, Yutong Ban

November 2026

A simulated arm with a dexterous hand pulls a kitchen drawer open.

TL;DR DexArena is a SAPIEN benchmark of 10 long-horizon dexterous-hand tasks split into 30 stages, with 3,000 replay-verified VR teleoperation demos. The best baselines solve about 60% of single stages, but chained end to end only ACT ever finishes a full task (1.2% on average).

What they do

DexArena is a simulation benchmark, built on SAPIEN, for testing whether dexterous-hand policies can learn single skills, generalize them to new scenes and chain them into long tasks. It has 10 long-horizon tasks, such as juice preparation or furniture assembly, each split into 2-5 stages, for 30 stage-level tasks in six motion categories.

How it works

  • Shared stages. A stage task uses the same layout, physics, randomization and success test as its slice of the long task. Success is a threshold on joint angles, relative poses or insertion depth that must hold for 3-40 steps.
  • Embodiments. One interface covers a parallel-jaw gripper and three hands (Sharpa, Wuji, XHand), single-arm or bimanual.
  • Data. VR teleoperation (Quest 3, MANUS gloves) at 20 Hz. Every frame stores the scene state, so episodes can be re-rendered from new cameras or scenes. 100 demos per stage that succeed on open-loop replay: 3,000 episodes, 9.1 h.
  • Baselines. PPO and RLPD with demonstrations, ACT, Diffusion Policy, GR00T and π0.5, with 50 evaluation episodes per task.

Results

  • Average stage success: GR00T 61.4%, ACT 58.0%, RLPD 54.3%, π0.5 52.1%, DP 39.2%, PPO 27.9%.
  • In three held-out scenes, π0.5 (49.2%) and ACT (45.4%) hold up best; RLPD falls to 17.8% and DP to 17.2%.
  • Chained end to end, only ACT completes any full task (8% furniture assembly, 4% package opening, 1.2% average).
  • Open-loop replay of the accepted demos succeeds 98% of the time for opening and closing but only 65% for insertion and assembly.

Limitations

Simulation only. The authors calibrated the simulator but say multi-finger contact still leaves a non-negligible sim-to-real gap. Baselines are not compute-matched, and most of π0.5’s long-horizon entries were still awaiting final validation.

Videos

The dexterous hand opens a laptop lid on the desk.
The hand picks up a tool to open a package on the table.

Figures

From Figure 1: DexArena overview, with 10 long-horizon tasks, 30 stage-level tasks, 3,000 teleoperated episodes, seve...
From Figure 1: DexArena overview, with 10 long-horizon tasks, 30 stage-level tasks, 3,000 teleoperated episodes, several dexterous embodiments and household and industrial scenes.
From Figure 2: The ten long-horizon tasks and the stages each one is split into.
From Figure 2: The ten long-horizon tasks and the stages each one is split into.
From Figure 3: Data collection. The operator wears a VR headset and gloves with a controller on the back of each hand...
From Figure 3: Data collection. The operator wears a VR headset and gloves with a controller on the back of each hand; finger and wrist inputs become joint targets, and each episode is reviewed, validated and rendered offline.
From Figure 6: The training scene and the three held-out kitchen scenes used to test visual generalization.
From Figure 6: The training scene and the three held-out kitchen scenes used to test visual generalization.
Abstract

Long-horizon dexterous manipulation remains a fundamental challenge for robot learning, requiring the composition of multiple skills, precise bimanual coordination, generalizable visual understanding, and robust skill transitions. Yet existing manipulation benchmarks largely emphasize isolated skills or support only a limited range of end-effectors, making these capabilities difficult to evaluate systematically. We introduce DexArena, a benchmark designed to evaluate learning and generalization in long-horizon bimanual dexterous manipulation. DexArena features 10 long-horizon tasks and 30 stage-level tasks spanning diverse household and industrial scenarios. Each long-horizon task is composed of explicitly defined stages with physically grounded success conditions, enabling reliable evaluation of both intermediate progress and end-to-end task completion. A unified benchmark interface and verified human-teleoperated demonstrations enable testing different policy learning algorithms, with standardized simulation configurations, scene randomization, trajectory replay, and task-level evaluation. Through extensive baseline experiments, we identify substantial challenges in precise manipulation, generalization to scene variation, and skill composition. DexArena establishes a challenging and reproducible benchmark for developing robot learning methods that move toward generalizable long-horizon dexterous behavior.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.