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

View the PDF file for the paper entitled Robotwin 2.0: Development Data generator and standard with random distribution a strong field for powerful automated manipulation, by Tianxing Chen and 25 other books

PDF HTML (experimental) view

a summary:The synthesis of simulation -based data has emerged as a strong model to enhance automated manipulation in the real world. However, current data groups are still insufficient to manipulate strong (1) that there are no methods of generating developmental tasks and (2) excessive simulation environments in simplification. We offer Robotwin 2.0, which is a developed framework for a wide -ranging automatic generation of diverse and realistic data, along with uniform evaluation protocols for dual arm manipulation. In essence, Robotwin-De, a library of 731 cases across 147 categories with semantic and manipulation slices. Accordingly, we design a pipeline to synthesize expert data that benefits from multimedia language models (MLMS) and improves simulation in the episode to create implementation code at the task level automatically. To improve the transfer of SIM to Real, Robotwin 2.0 applies the random distribution of the organized field along five axes: chaos, lighting, background, table height, language, and promoting diversity of data policy meal. A frame is created across 50 double weapons and five robot embodiment. Experimental, it results in an increase of 10.9 % in the success rate of the code generation. To learn the estuary policy, the VLA model achieves the artificial data in addition to only 10 real demonstrations relatively improved by 367 % on the 10 -DO foundation, while the zero models that were trained only on artificial data get a 228 % profit. These results shed light on the effectiveness of Robotwin 2.0 in promoting the transfer of SIM to real and powerful for environmental changes. We issue a data generator, standard, data set, and symbol to support developmentable research in strong manipulation. Project page: URL https, this symbol: URL https.