Per-step Integrated Gradients analysis for diffusion-policy robotics models
The repo offers a full pipeline to compute per-step Integrated Gradients for Robotics Diffusion Transformer (RDT) models, reproducing the paper's metrics on ManiSkill3 manipulation tasks. It includes scripts, notebooks, and released per-step metric records for faithfulness, sanity, and displacement evaluations. Researchers and robotics developers can use it to attribute decisions to vision, language, and state inputs, and to benchmark attribution methods. Compared to ad‑hoc analysis, it provides a reproducible, end‑to‑end framework with ready‑to‑run evaluation utilities.
View on GitHub →itscool2b/ig-work-2026