This page organizes papers related to artificial intelligence published around the world. This page is summarized using Google Gemini and is operated on a non-profit basis. The copyright of the paper belongs to the author and the relevant institution. When sharing, simply cite the source.
CUPID: Curating Data your Robot Loves with Influence Functions
Created by
Haebom
Author
Christopher Agia, Rohan Sinha, Jingyun Yang, Rika Antonova, Marco Pavone, Haruki Nishimura, Masha Itkina, Jeannette Bohg
Outline
This paper highlights that policy performance in robot imitation learning is highly dependent on the quality and composition of demonstration data, yet it is challenging to precisely understand how individual demonstrations contribute to outcomes such as closed-loop task success or failure. Therefore, we propose CUPID, a robot data curation method based on a novel influence function theoretical formulation for imitation learning policies. CUPID estimates the impact of each training demonstration on the expected return of a policy by considering a set of evaluation rollouts, allowing it to rank and select demonstrations based on their impact on the policy's closed-loop performance. CUPID is used to curate data by filtering out training demonstrations detrimental to policy performance and subselecting novel trajectories that most likely improve the policy. Simulation and hardware experiments demonstrate that the method consistently identifies data that drives performance at test time. For example, state-of-the-art diffusion policies can be achieved on simulated RoboMimic benchmarks by training with less than 33% of the curated data, and similar performance improvements are observed on hardware. Furthermore, hardware experiments demonstrate that it can identify strategies robust to distributional shifts, isolate spurious correlations, and even improve the post-training performance of common robot policies. Code and video are available at https://cupid-curation.github.io .
