Reproduce “TapSampling: Inference-Time Sampling with a Task-Progress-Understanding Verifier for Robotic Manipulation” on OpenResearch

About this paper

Sizhe Zhao, Shengping Zhang, Shuo Yang, Weiyu Zhao, Shuigen Wang, Xiangyang Ji

TapSampling, an inference-time framework from Harbin Institute of Technology and Tsinghua University, enhances robotic manipulation by employing efficient action sampling via a learned posterior and an interpretable verifier that understands task progress. The system improved average success length on CALVIN benchmarks by up to 0.21, increased average success rates on LIBERO-Long by 1.2%, and achieved a 5.0% average success rate improvement in real-world manipulation tasks.

Key insights

• Inference-time computation serves as an effective complementary axis for improving robotic policy performance, offering benefits beyond solely scaling training data or model size.
• A learned posterior distribution, modeled by an Action-VAE, can efficiently generate high-fidelity and diverse action candidates that approximate the policy's distribution, addressing the trade-off between speed and fidelity.
• Task progress can be self-supervisedly extracted from sequential robotic datasets and used as an interpretable metric for verifying actions, distinguishing beneficial from detrimental low-level robot behaviors.

Summary by alphaXiv.