Description

Contact-rich manipulation requires adapting stiffness to uncertain dynamics. While Reinforcement Learning (RL) can learn these Variable Impedance policies, it suffers from unsafe exploration and poor sample efficiency. This thesis proposes an Adaptive Variable Impedance Controller guided by Multimodal Safety Priors (Vision, Language, Proprioception, Forces). This approach constrains RL exploration to ensure safe, data-efficient learning of complex interactions.

The work is under the scope of the following projects:

• DROVIC: Data-driven Robot Variable Impedance Control with Context Awareness (web)