This research proposal addresses the 'seeing-to-doing gap' in embodied AI by introducing Morphology-Aware Action Representation (MAAR), a framework enabling zero-shot transfer of manipulation policies across heterogeneous robot morphologies. The method transforms robot-specific actions into a unified task-space representation (MAUS) using morphology encoders based on graph neural networks processing URDF structures. A cross-embodiment policy conditioned on morphology embeddings generates actions transferable to novel robots without target-domain training data. The proposal outlines comprehensive experiments spanning simulation (ManiSkill3, RLBench) and real-world validation (Franka, xArm), with expected outcomes demonstrating >50% zero-shot success rates.
Key findings
Morphology-Aware Unified Space (MAUS) normalizes actions to task-space (SE(3)) enabling representation across different kinematic structures
Graph-based morphology encoder processes URDF structures to generate embeddings that condition the cross-embodiment policy
Zero-shot transfer mechanism allows direct deployment on target robots without training data by decoding task-space actions through target morphology-specific transformations
Proposed experimental matrix validates transfer across 7-DOF and 6-DOF arms with ablation studies confirming component criticality
Risk analysis identifies sim2real gap and few-shot adaptation as contingency strategies if zero-shot performance falls below 30%
Limitations & open questions
Document presents a research proposal without empirical validation or experimental results to support the zero-shot transfer hypothesis
Relies on speculative future-dated citations (Ai et al., 2025) regarding embodiment scaling laws that are not yet established in literature
Assumes task-space representation suffices for all manipulation tasks, potentially limiting applicability to complex contact-rich scenarios requiring detailed torque control