This is a page that curates AI-related papers published worldwide. All content here is summarized using Google Gemini and operated on a non-profit basis. Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.
RoboTwin 2.0 is a large-scale, diverse, and realistic data generation framework for scalable dual-arm manipulation. To overcome the limitations of existing datasets (lack of scalable task generation methods and oversimplified simulation environments), we designed an expert data synthesis pipeline utilizing a multimodal language model (MLLM) and simulation-based refinement based on the RoboTwin-OD object library, which contains 731 object instances (147 categories). We applied structured domain randomization across five axes (clutter, lighting, background, table height, and language) to improve simulation-to-reality transfer and enhance data diversity and policy robustness. Applying this framework to 50 dual-arm tasks and five robot models, we achieved a 10.9% improvement in code generation success rate, a 367% relative performance improvement when training a VLA model using synthetic data and 10 real-world demos, and a 228% performance improvement over a zero-shot model trained solely on synthetic data. We support scalable, robust dual-arm manipulation research by releasing data generators, benchmarks, datasets, and code.
Takeaways, Limitations
•
Takeaways:
◦
Providing a large-scale, diverse, and realistic synthetic data generation framework for scalable dual-arm manipulation.
◦
An efficient task generation pipeline is presented using a multimodal language model and simulation-based improvements.
◦
Improving simulation-to-real transition performance and ensuring robustness to environmental changes through structured domain randomization.
◦
Effective policy learning and zero-shot performance improvement using synthetic data.
◦
Providing research sharing and scalability through data generators, benchmarks, datasets, and code disclosure.
•
Limitations:
◦
The variety of robot models and tasks currently supported may be limited.
◦
It is difficult to achieve a perfect match with the real environment, so additional adjustments may be required when applying to the real environment.
◦
The quality of data generation may be affected by the performance of MLLM.
◦
The scope of structured domain randomization needs to be further expanded.
