Reproduce “ROPA: Synthetic Robot Pose Generation for RGB-D Bimanual Data Augmentation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jason Chen, I-Chun Arthur Liu, Gaurav Sukhatme, Daniel Seita
Researchers at the University of Southern California developed ROPA, an offline data augmentation method that generates synthetic robot poses and corresponding action labels for RGB-D bimanual imitation learning. This approach consistently improved policy success rates and robustness across various simulated and real-world tasks, outperforming baseline methods by enabling the synthesis of diverse training data.
Key insights
- Adapting pose-guided image synthesis with diffusion models can effectively generate diverse robot poses and corresponding action labels for bimanual manipulation.
- Visually rich skeleton representations, rather than raw numerical joint data, serve as more effective conditioning inputs for diffusion models in robot image synthesis.
- Incorporating constrained optimization during action label generation is crucial for ensuring the physical feasibility and consistency of augmented bimanual robot poses, especially in contact scenarios.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.