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.
Empirical Analysis of Sim-and-Real Cotraining of Diffusion Policies for Planar Pushing from Pixels
Created by
Haebom
Author
Adam Wei, Abhinav Agarwal, Boyuan Chen, Rohan Bosworth, Nicholas Pfaff, Russ Tedrake
Outline
Cooperative learning using demonstration data generated from simulations and real hardware has emerged as a promising approach to scale imitation learning in robotics. This study aims to elucidate the fundamental principles of this simulation-real co-learning approach to inform simulation design, simulation and real data set generation, and policy training. Experimental results demonstrate that co-learning using simulated data can significantly improve performance, especially when real data is limited. This performance improvement scales with additional simulation data until a plateau is reached, and the performance ceiling increases with the addition of more real data. Furthermore, for non-grasping or contact-intensive tasks, reducing the physical domain gap may be more effective than increasing visual fidelity. Somewhat surprisingly, we find that some degree of visual disparity can be beneficial for co-learning. Binary probes demonstrate that high-performing policies must learn to distinguish between simulated and real domains. We conclude by investigating the subtle differences and mechanisms that facilitate positive transfer between simulation and real-world tasks. Our narrow focus on the common task of planar pushing allowed us to conduct a thorough study. We conducted experiments involving over 50 real-world policies (evaluated over 1,000 trials) and 250 simulated policies (evaluated over 50,000 trials). The video and code can be found at https://sim-and-real-cotraining.github.io/ .
Cooperative learning using simulated data can significantly improve performance in robot-imitation learning where real-world data is limited.
◦
While additional simulation data contributes to performance improvements, the effect diminishes beyond a certain point. Adding real-world data increases the performance ceiling.
◦
For tasks that involve a lot of hand-holding or contact, reducing the physical domain gap may be more important than visual fidelity.
◦
A certain level of visual differentiation can aid collaborative learning by improving the ability to distinguish between simulated and real domains.
•
Limitations:
◦
The study was limited to a specific task, flat pushing, and further research is needed to determine its generalizability.
◦
Further research is needed to determine optimal ways to bridge the domain gap between simulations and real-world data.
◦
Further research is needed on scalability to various robotic tasks and environments.
