Reproduce “Policy Contrastive Decoding for Robotic Foundation Models” 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
Shihan Wu, Xu Luo, Ji Zhang, Junlin Xie, Jingkuan Song, Heng Tao Shen, Lianli Gao
Policy Contrastive Decoding (PCD), a training-free method, addresses spurious correlations in robotic foundation models by contrasting action probability distributions from original and object-masked inputs. This approach enhanced task success rates by up to 50.6% in simulation and 108% in real-world experiments for various baseline policies.
Key insights
- Spurious correlations in robot policies can be mitigated during inference by combining action probability distributions from original and object-masked inputs, thereby amplifying object-relevant feature reliance.
- Existing generalist robot policies can be treated as black-box models, enabling a training-free, plug-and-play solution through a pipeline of off-the-shelf vision models for input manipulation.
- Extending contrastive decoding to diffusion-based policies is achievable by approximating their action probability distributions using Kernel Density Estimation (KDE) from sampled actions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.