AI Paper Atlas

Research map / Vision-language and multimodal / Robot action and navigation

RoboTwin 2.0: A Scalable Data Generator and Benchmark with Strong Domain Randomization for Robust Bimanual Robotic Manipulation

Tianxing Chen, Zanxin Chen, Baijun Chen, Zijian Cai, Yibin Liu, Zixuan Li, Qiwei Liang, Xianliang Lin, Yiheng Ge, Zhenyu Gu, Weiliang Deng, Yubin Guo and 14 more

ICML 2026 · Vision-language and multimodal › Robot action and navigation · approach: Memory and subgoal-based manipulation

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add multimodal VLM feedback to code generation (ablation)RoboTwin 2.0 code agent · 10 code-gen tasksASR: 62.1% -> 71.3%Improved
add multimodal VLM feedback to RoboTwin 1.0 code generation (ablation)RoboTwin 1.0 code agent · 10 code-gen tasksASR: 47.4% -> 63.9%Improved
fine-tune VLA on clean non-randomized synthetic data (ablation)RDT and Pi0 · 8 sim tasks, domain-randomized evalsuccess rate: RDT 18.8% -> 14.6%; Pi0 22.5% -> 24.9%No clear effect
pretrain on domain-randomized RoboTwin 2.0 data (ablation)RDT · 8 sim tasks, domain-randomized evalsuccess rate: 18.8% -> 24.8%Improved
pretrain on domain-randomized RoboTwin 2.0 data (ablation)Pi0 · 8 sim tasks, domain-randomized evalsuccess rate: 22.5% -> 29.1%Improved
add 1k domain-randomized synthetic trajectories to 10 real demos (ablation)RDT · real-world, unseen background, clutteredsuccess rate: 9.0% -> 42.0%Improved
train on 1k synthetic trajectories only (zero-shot) (ablation)RDT · real-world, unseen backgroundssuccess rate: 15.5% -> 36.5% (clean); 9.0% -> 29.5% (cluttered)Improved
add 1k synthetic data to real demos on Click Bell, seen clean (ablation)RDT · real-world Click Bell, seen, unclutteredsuccess rate: 36.0% -> 24.0%Hurt
add embodiment-aware grasp adaptation (RoboTwin 2.0 vs 1.0) (ablation)data generation pipeline · 50 tasks, 5 embodimentsdata collection success rate: +8.3% average; Piper +22.7%Improved
add embodiment-aware grasp adaptation on high-DoF arms (ablation)data generation pipeline · Franka, UR5data collection success rateNo clear effect
evaluate on domain-randomized (Hard) instead of clean (Easy) conditionsRDT and Pi0 · RoboTwin 2.0 benchmarksuccess rate: -20.8% RDT; -30.1% Pi0Hurt
use VLA pretraining vs non-pretrained policies under Hard conditionsRDT, Pi0 · RoboTwin 2.0 benchmark Hardaverage success rate: Pi0 16.3, RDT 13.7 vs ACT 1.7, DP 0.6Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy