Graph Based Morphology Encoding for Robust Generalisation of RL Policies
RL policies need to be re-trained and fine tuned everytime when being tested on a new design. This research focusses on encoding methods to assist in application of generalised policies over a new real world robot design in simulation
Research Problem
TThis research addresses the generalization of reinforcement learning (RL) policies across different robot morphologies, with the goal of reducing retraining when robot designs change.
• Problem: Existing RL policies must be retrained for new robot designs, limiting scalability and real-world deployment.
• Current Limitation: Transformer-based approaches such as MetaMorph enable cross-morphology training but do not explicitly encode structural connectivity, leading to weak transfer on complex, heterogeneous robots like the Unitree G1.
• Research Gap: Morphological encodings insufficiently capture graph-level relationships between links and joints, especially in underactuated or non-uniform real-world robots.
• Proposed Approach: Model robot URDFs as graphs and integrate graph-based encodings (GNN/GAT/GCNT) into Transformer actor–critic architectures trained with PPO.
• Objective: Improve structural awareness and inter-module communication to enable strong zero-shot generalization to unseen robot designs in simulation.