Reproduce “Context-Aware Force Estimation for Deformable Tool Manipulation in Robotic Environmental Swabbing via Few-Shot Continual Adaptation” on OpenResearch

About this paper

Siavash Mahmoudi, Chaitainya Kuppar Reddy, Yang Tian, Dongyi Wang

A data-driven framework estimates tip-level contact forces during robotic deformable tool manipulation (DTM) for environmental swabbing using only proprioceptive sensing. It employs a context-aware, few-shot continual adaptation strategy that achieved 18-63% error reduction on new surfaces/tools while preventing catastrophic forgetting.

Key insights

• Recurrent architectures like LSTMs are highly effective for modeling the non-Markovian, history-dependent dynamics of deformable tool-environment interactions, outperforming Transformer and TCN models.
• Low-dimensional latent context vectors, modulated via FiLM, can effectively encode task-specific dynamics for rapid and data-efficient adaptation to novel interaction conditions.
• Freezing the shared recurrent backbone and exclusively optimizing context embeddings during adaptation successfully prevents catastrophic forgetting, enabling lifelong learning across diverse interaction regimes.

Summary by alphaXiv.