DexArenaDexterous intelligence
Open drawer
Open laptop
Insert rail
Long-horizon dexterous manipulation

DexArena

Dexterity over long horizons.

Execute precise skills, generalize across scenes, and compose behavior into complete bimanual tasks.

Discover the research↗
DexArena cabinet task rendered in a RoboCasa scene
Cross-scene evaluation · 90 combinations

Same skill.
New scene.

Change the scene. Keep the task.

Evaluate all 30 canonical stages on three native scene surfaces with fixed policy cameras.

Explore generalization↗
Two dexterous hands opening a package and retrieving an apple in a native kitchen scene
Featured task · Bimanual dexterity

Two hands.
One task.

Open the package. Retrieve the apple.

Coordinate both hands to open the flaps, then grasp and place the apple: two connected stages with physical success conditions.

Explore Package Opening↗
The research behind DexArenaPaper ↗

DexArena: A Unified Benchmark for Long-Horizon Dexterous Manipulation

Authors

  1. Haonan Ping3
  2. Jianshu Hu1
  3. Yimin Wang3
  4. Yi Zhu4
  5. Pai Liang5
  6. Qiyan Jin6
  7. Qirui Bao1
  8. Ying Qian7
  9. Xinyang Ding1
  10. Hongfei Qiao1
  11. Tianle Xue1
  12. Yucheng Bai1
  13. Soo Wei Jie1
  14. Shuguang Li2
  15. Yutong Ban1

Affiliations

  8. PAIR Shanghai Institute for Physical Artificial Intelligence and Robotics
Paper
Paper ↗

Not publicly available yet.

Code
Code Coming Soon
Leaderboard
View benchmark results →
01 · Benchmark design

Skills and sequences in a shared physical setting.

Ten complete tasks and thirty matched stage environments connect household, laboratory and industrial interactions. Persistent physical predicates measure both intermediate progress and complete behavior.

Figure 1: DexArena overview of tasks, objects, embodiments, skills and scenarios
Figure 1 · DexArena overview. Multi-stage tasks, diverse scenarios, teleoperated demonstrations and multiple embodiments. Full resolution ↗
02 · Demonstrations

From human motion to replayable demonstrations.

VR wrist tracking and glove inputs are mapped into absolute arm and hand targets. Episodes then pass through review, export and replay validation, with recorded simulator states retained for offline multi-view rendering.

Figure 3: Data collection pipeline showing VR and glove input, motion mapping, simulation feedback, episode review and replay validation, and offline RGB rendering
Figure 3 · Data collection pipeline. The operator wears a VR headset and gloves, with a controller mounted on each hand. Finger and wrist inputs are mapped to absolute joint targets, while VR rendering provides scene feedback. Episodes are recorded, reviewed, exported, validated and rendered offline. Full resolution ↗
3,000

Replay-successful demonstrations

100 episodes per canonical stage, recorded at 20 Hz.

State + action

Restoration and replay

Recorded state supports fidelity checks and open-loop trajectory validation.

Multi-view

Offline RGB rendering

The same recorded interaction can be rendered from fixed base and wrist views.

03 · Findings

Strong skills are not yet reliable sequences.

Six baselines reveal distinct limitations in stage execution, held-out-scene generalization and end-to-end task composition. Compare baseline results with shared-scale charts, heatmaps and downloadable tables.

01

Stage execution

30 stages · training scene

Compare policies →02

Visual generalization

30 stages × 3 native scenes

Explore scene transfer →03

Long-horizon composition

10 sequences · no stage resets

Inspect sequence success →

These are research-preview baseline results, not community submissions.

Simulation tasks

Ten sequences. Thirty measurable stages.

From kitchen preparation to precision assembly, each long-horizon task is decomposed into physically grounded stages.

Meal Preparation Cabinet & Pouring Drawer & Plug Nut & Bolt Office Desk
Furniture Assembly Pipette Transfer Juice Preparation Package Opening Table Cleaning
View all task definitions↗
Benchmark architecture

Every stage has a place
in a larger task.

Matched environments connect isolated skills to complete sequences. The same definitions carry through data collection and cross-scene evaluation.

Long-horizon tasks
10

Complete, dependent sequences

Canonical stage tasks
30

Physically grounded success conditions

Sequence → stage structure2–5 stages per task
01020304050607080910

Each column is one task. Each segment is one measurable stage.

Explore the task definitions
Replay-validated episodes
3,000
100 per stage
× 30 stages

Human demonstrations selected after successful action replay, with simulator states retained for restoration.

9.1 h of demonstrations20 Hz recording
Stage–scene combinations
90

30 stages. 3 native kitchens.
Fixed tasks, different surroundings.

Blue stone
Light wood
Warm wood
Each mark = one stage–scene pair
Benchmark design

Measure the transition from a skill to a behavior.

Pipette task in native kitchen scene 2Pipette task in native kitchen scene 1Pipette task in native kitchen scene 6
02

Visual generalization

Keep task definitions fixed while changing the native kitchen environment.

Bimanual package-opening task
Embodiment

29 dimensions with one hand. 58 with two.

The released canonical data uses Panda arms with 22-DoF Sharpa hands and absolute joint-position targets: 7 arm + 22 hand joints per side.

  • Single-arm and bimanual variants
  • Shared task and success interfaces
  • Parallel-jaw, Sharpa, Wuji, and XHand simulation support
Human demonstrations

Recorded, reviewed, restored, replayed.

The data pipeline retains the simulator state needed for deterministic restoration and offline multi-view rendering.

  1. MANUS Metagloves Pro, official manufacturer product imageMANUS · Official product image ↗
    01 · Capture

    Recorded

    Human wrist and finger motion becomes robot joint targets at 20 Hz.

  2. A lunch-box opening demonstration for review
    02 · Quality check

    Reviewed

    Review the episode and retain demonstrations that complete the task.

  3. Robot and pipette objects restored in a native scene
    03 · State restoration

    Restored

    Recover saved simulator states for offline multi-view rendering.

  4. 04 · Action replay

    Replayed

    Replay recorded actions and select 100 successful episodes per stage.

3,000 demonstrations · 9.1 h · 20 Hz
Generalization

One task. Three native scenes.

The same pipette workstation in blue stone, light wood and warm wood kitchens. Evaluate all 30 stages across these three native backgrounds.

30 × 390 task–scene combinations
Blue kitchen: the same pipette task with unchanged object layout

Blue kitchen

Stone islandrobocasa_l3_s2
Light wood kitchen: the same pipette task with unchanged object layout

Light wood kitchen

Wood islandrobocasa_l3_s1
Warm wood kitchen: the same pipette task with unchanged object layout

Warm wood kitchen

Stone islandrobocasa_l3_s6

Fixed policy cameras and shared task definitions. Optional LabUtopia laboratory for further studies.

Experimental results

Learning a skill is only the beginning.

Six baselines evaluated on execution, visual generalization and long-horizon composition. All scores below are reported in the supplied manuscript.

01 / 30 stages · training scene

Strong individual skills.

GR00T records the highest average success in the training scene. ACT and RLPD follow, but no method consistently solves every stage.

+3.4 ppadvantage over ACT
Stage-level results
Average success by policySuccess (%) · 0–100
  1. GR00T61.4 percent success
  2. ACT58.0 percent success
  3. RLPD54.3 percent success
  4. π0.552.1 percent success
  5. Diffusion Policy39.2 percent success
  6. PPO27.9 percent success

50 evaluation episodes per stage. Scores are averaged over 30 canonical stages.

02 / 30 stages × 3 native scenes

A different scene changes the ranking.

π0.5 leads across three held-out kitchens. Its average is 2.9 percentage points below the training-scene result; the same policy is evaluated without adaptation.

+3.8 ppadvantage over ACT
Cross-scene results
Average success by policySuccess (%) · 0–100
  1. π0.549.2 ± 2.3 percent success
  2. ACT45.4 ± 3.6 percent success
  3. GR00T41.7 ± 14.2 percent success
  4. PPO21.7 ± 3.5 percent success
  5. RLPD17.8 ± 10.9 percent success
  6. Diffusion Policy17.2 ± 5.7 percent success

50 episodes per stage per scene. ± denotes standard deviation across three scene averages, not a confidence interval.

03 / 10 sequences · no stage resets

The sequence remains the challenge.

π0.5 and ACT report non-zero average success on complete sequences. Successful individual stages do not yet translate into reliable end-to-end behavior.

2 / 6methods with non-zero sequence success
Complete-sequence results
Average success by policySuccess (%) · 0–100
  1. π0.51.6 percent success
  2. ACT1.2 percent success
  3. PPO0.0 percent success
  4. RLPD0.0 percent success
  5. Diffusion Policy0.0 percent success
  6. GR00T0.0 percent success

50 episodes per complete task. Success requires reaching the final task condition without resets between stages.

Manuscript baseline results · 6 methods · Not a live submission rankingHow to read these results ↗
Paper baselines

Explore every result.

Compare PPO, RLPD, ACT, Diffusion Policy, GR00T and π0.5 across all stages and complete sequences.

Open leaderboard↗