A holistic understanding of object properties across diverse sensory modalities (e.g., visual, audio, and haptic) is essential for tasks ranging from object categorization to complex manipulation. Drawing inspiration from cognitive science studies that emphasize the significance of multi-sensory integration in human perception, we introduce MOSAIC (Multimodal Object property learning with Self-Attention and Interactive Comprehension), a novel framework designed to facilitate the learning of unified multi-sensory object property representations. While it is undeniable that visual information plays a prominent role, we acknowledge that many fundamental object properties extend beyond the visual domain to encompass attributes like texture, mass distribution, or sounds, which significantly influence how we interact with objects. In MOSAIC, we leverage this profound insight by distilling knowledge from multimodal foundation models and aligning these representations not only across vision but also haptic and auditory sensory modalities. Through extensive experiments on a dataset where a humanoid robot interacts with 100 objects across 10 exploratory behaviors, we demonstrate the versatility of MOSAIC in two task families: object categorization and object-fetching tasks. Our results underscore the efficacy of MOSAIC's unified representations, showing competitive performance in category recognition through a simple linear probe setup and excelling in the fetch object task under zero-shot transfer conditions. This work pioneers the application of sensory grounding in foundation models for robotics, promising a significant leap in multi-sensory perception capabilities for autonomous systems. We have released the code, datasets, and additional results: https://github.com/gtatiya/MOSAIC.

Humans leverage multiple sensor modalities when interacting with objects and discovering their intrinsic properties. Using the visual modality alone is insufficient for deriving intuition behind object properties (e.g., which of two boxes is heavier), making it essential to consider non-visual modalities as well, such as the tactile and auditory. Whereas robots may leverage various modalities to obtain object property understanding via learned exploratory interactions with objects (e.g., grasping, lifting, and shaking behaviors), challenges remain: the implicit knowledge acquired by one robot via object exploration cannot be directly leveraged by another robot with different morphology, because the sensor models, observed data distributions, and interaction capabilities are different across these different robot configurations. To avoid the costly process of learning interactive object perception tasks from scratch, we propose a multi-stage projection framework for each new robot for transferring implicit knowledge of object properties across heterogeneous robot morphologies. We evaluate our approach on the object-property recognition and object-identity recognition tasks, using a dataset containing two heterogeneous robots that perform 7,600 object interactions. Results indicate that knowledge can be transferred across robots, such that a newly-deployed robot can bootstrap its recognition models without exhaustively exploring all objects. We also propose a data augmentation technique and show that this technique improves the generalization of models. We release code, datasets, and additional results, here: https://github.com/gtatiya/Implicit-Knowledge-Transfer.

Humans learn about object properties using multiple modes of perception. Recent advances show that robots can use non-visual sensory modalities (i.e., haptic and tactile sensory data) coupled with exploratory behaviors (i.e., grasping, lifting, pushing, dropping, etc.) for learning objects' properties such as shape, weight, material and affordances. However, non-visual sensory representations cannot be easily transferred from one robot to another, as different robots have different bodies and sensors. Therefore, each robot needs to learn its task-specific sensory models from scratch. To address this challenge, we propose a framework for knowledge transfer using kernel manifold alignment (KEMA) that enables source robots to transfer haptic knowledge about objects to a target robot. The idea behind our approach is to learn a common latent space from multiple robots' feature spaces produced by respective sensory data while interacting with objects. To test the method, we used a dataset in which 3 simulated robots interacted with 25 objects and showed that our framework speeds up haptic object recognition and allows novel object recognition.

When identifying an object and its properties, humans use features from multiple sensory modalities produced when manipulating the object. Motivated by this cognitive process, we propose a deep learning methodology for object category recognition which uses visual, auditory, and haptic sensory data coupled with exploratory behaviors (e.g., grasping, lifting, pushing, etc.). In our method, as the robot performs an action on an object, it uses a Tensor-Train Gated Recurrent Unit network to process its visual data, and Convolutional Neural Networks to process haptic and auditory data. We propose a novel strategy to train a single neural network that inputs video, audio and haptic data, and demonstrate that its performance is better than separate neural networks for each sensory modality. The proposed method was evaluated on a dataset in which the robot explored 100 different objects, each belonging to one of 20 categories. While the visual information was the dominant modality for most categories, adding the additional haptic and auditory networks further improves the robot's category recognition accuracy. For some of the behaviors, our approach outperforms the previous published baseline for the dataset which used handcrafted features for each modality. We also show that a robot does not need the sensory data from the entire interaction, but instead can make a good prediction early on during behavior execution.