RoboTwin 2.0 Leaderboard

Benchmark Setting

Training data is fixed to 50 demo_clean trajectories × 50 tasks (2,500 demos total) on Aloha-AgileX, then evaluated 100 trials/task under demo_clean and demo_randomized. Use the ranking control to switch titles between Average（c2c+c2r）, clean2random（hard）, and clean2clean（easy）. Co-train and Single-task SFT share one board. The default ranking is Average（c2c+c2r）.

Submission & listing policy

To be listed on this leaderboard, a model must provide publicly released code, publicly released weights, and a technical report (arXiv paper or equivalent public document) describing the method. To submit results or request listing, contact chentianxing2002@gmail.com with links to the code, weights, and report.

XPolicyLab

A Unified Standard and Open Ecosystem for Robot Policy Evaluation and Deployment

Institutions MMLab@HKU & THU
Paper arXiv:2608.09892
Code github.com/XPolicyLab/XPolicyLab
Contact Tianxing Chen (project lead), chentianxing2002@gmail.com
Website xpolicylab.github.io

The same adapters serve this RoboTwin 2.0 board and RoboDojo (sim + real). Results here are reproduced via the XPolicyLab standard interface.

Data: 50 demo_clean × 50 tasks Co-train: one policy jointly trained on all 50 tasks Single: one checkpoint fine-tuned per task (paper baselines) Average（c2c+c2r）: ranked by (c2c + c2r) / 2 clean2random（hard）: ranked by demo_randomized clean2clean（easy）: ranked by demo_clean Open: public code + weights + technical report required XPolicyLab: bridges RoboTwin ↔ RoboDojo Contributor: RoboTwin Team
Ranking setting
View
Ranked by Average（c2c+c2r） mean ↓
Current ranking setting

Average（c2c+c2r）

Train: 50 demo_clean × 50 tasks (2,500) Eval / sort: (c2c + c2r) / 2 Robot: Aloha-AgileX Track: Co-train (+ Single SFT baselines)

Overall Ranking — Average（c2c+c2r）

Latest update:
Rank Method Contributor Average（c2c+c2r） clean2random（hard） clean2clean（easy） Track Date

Per-task Results

Columns follow the current Average（c2c+c2r） / clean2random（hard） / clean2clean（easy） ranking. Methods tagged Single are single-task SFT baselines.

RoboTwin XPolicyLab RoboDojo

XPolicyLab bridges RoboTwin and RoboDojo with one policy interface, data format, and train/eval pipeline. Website · Code

Sim Leaderboard Top 10 · Ranked by Average · Score ↓

Dimension
Rank by
Official Leaderboard Updated:

XPolicyLab bridges RoboTwin and RoboDojo: one policy interface, one data format, and one train/eval pipeline. Models listed here can be transferred to RoboDojo (sim + real) without rewriting a line of code. Preview below shows the official Top 10; cells are Score / SR% per capability dimension.

Show full leaderboard on robodojo-benchmark.com

News