ManiMorph: Object Representations in Robot Manipulators Morphology for Improving Multi-Task Manipulation Performance
Ali Abdalla, Michael Przystupa, Xinrui Zu, Kevin Sebastian Luck, Glen Berseth
AI summary
Problem
Morphology-aware manipulation frameworks currently ignore how object interactions dynamically alter a robot's kinematic chain, limiting policy robustness and multi-task generalization.
Approach
ManiMorph unifies robot limbs and target objects into a single graph processed by a Transformer, using FiLM layers to condition the network on task-specific requirements without architectural changes.
Key results
- Object-as-node representation outperforms baselines on Lift and Door tasks
- FiLM task adapter enables robust multi-robot multi-task learning across control spaces
- Achieves zero-shot generalization to unseen object geometries and physical properties
- Surpasses alternative frameworks in cumulative reward and sustained contact control
Why it matters
Provides a scalable, morphology-aware foundation for robots to handle diverse objects and tasks without retraining, advancing general-purpose manipulation.
Abstract
Robot manipulation tasks involve direct interac- tions with objects, which can be viewed as dynamic changes to the robot’s kinematic chain. Morphology-aware learning frame- works, in which robot embodiment is explicitly modeled, do not account for these object-induced changes in their architectures. We address this gap by proposing ManiMorph, a multi-task, morphology-aware manipulation-learning framework in which object features are integrated into the robot’s morphological graph. We demonstrate that this node-centric representation, combined with a Feature-wise Linear Modulation (FiLM) task component, enhances the performance of the morphology-aware frameworks for robotic manipulation and generalizes effectively to new object variations.