Robotic manipulators are increasingly utilized in unstructured and novel environments. To interact with objects in the environment, it becomes critical to know their properties such as geometry, mass, surface friction coefficient, stiffness etc. Acquiring knowledge about these properties of novel objects could help in manipulating them stably and accurately, as well as predict the effects of various manipulation actions in advance. Perception of objects’ properties from visual and haptic sensing is a challenging problem, as these properties must be inferred indirectly from raw visual and tactile sensory information. In addition, often each property is only revealed under specific interactions, making it an interactive perception problem. To approach such a challenging problem, my research work takes inspiration from the human cognitive process involving interactive perception and active inference. A novel ‘predictive coding’ framework is developed under the scheme of Bayesian Inference. The key aspect of the framework is encoding interactions as the Probabilistic Markov Model along with learning schemes to learn the interaction model. Two key frameworks are developed – Active Differentiable Filter and Active Latent Filter to address the inference problem of diverse object properties like mass, center of mass, friction (characteristic) as well as texture, and softness (non-characteristic). Using the proposed approach, the robotic system I have set up can autonomously explore objects using raw and complementary vision and tactile information. The developed framework provides a step towards developing generalizable robotic manipulation skills and will be crucial to performing downstream robotic manipulation tasks efficiently.