Learning Lab LIVE: UNIT 12 – Intro to AXIOM and VBGS

This Learning Lab LIVE session introduces AXIOM and Variational Bayes Gaussian Splatting (VBGS) as two foundational technologies that bring Active Inference AI into practical, deployable systems. The session explains how these approaches move beyond the limitations of deep learning and reinforcement learning by enabling Intelligent Agents to learn, adapt, and coordinate in real time with far less data and computational overhead. Viewers are guided through Active Inference as a biologically grounded framework based on the Free Energy Principle, showing how prediction, action, and learning form a continuous loop.

The presentation then focuses on AXIOM, an open-source Active Inference framework developed by VERSES AI, designed for object-centric world modeling and adaptive decision-making. Viewers learn how AXIOM builds structured internal models of the world by identifying objects, tracking their identities, predicting their dynamics, and reasoning about interactions. This enables agents to generalize from minimal experience, plan under uncertainty, and continuously refine their understanding without retraining.

The session concludes with Variational Bayes Gaussian Splatting (VBGS), a probabilistic approach to real-time environmental modeling and navigation. VBGS allows agents to build and update uncertainty-aware 3D maps—or abstract belief maps for non-visual data—enabling safer exploration, smarter planning, and adaptive digital twins. Together with the Spatial Web Protocol, AXIOM and VBGS demonstrate how distributed, explainable, and energy-efficient intelligence can scale from individual agents to global enterprise systems.

Key Topics Covered

• Active Inference as intelligence in action
  Explains how Active Inference models learning as a continuous loop of prediction, observation, and action based on the Free Energy Principle.
• Why traditional AI struggles in real-world environments
  Examines the limitations of deep learning and reinforcement learning, including static training, poor generalization, and high data and energy requirements.
• Introduction to AXIOM
  Describes AXIOM as an open-source Active Inference framework for building agents that reason, learn, and coordinate in real time.
• Object-centric world modeling
  Explains how AXIOM represents environments as collections of objects rather than raw pixels, enabling faster learning and better generalization.
• Core priors and human-like intuition
  Shows how built-in priors allow AXIOM to rapidly infer object behavior, motion, and interactions with minimal experience.
• Probabilistic state-space models
  Explains how AXIOM represents possible world states, transitions, and beliefs using Bayesian inference.
• Slots and object hypotheses
  Introduces slots as internal representations of individual objects, each tracking position, velocity, identity, and other latent variables.
• Slot Mixture Model (SMM)
  Explains how AXIOM parses visual scenes into distinct objects by assigning pixels to object slots.
• Identity Mixture Model (IMM)
  Describes how AXIOM maintains consistent object identity even as appearance or position changes.
• Transition Mixture Model (TMM)
  Explains how AXIOM learns reusable motion patterns and predicts object dynamics across environments.
• Recurrent Mixture Model (RMM)
  Shows how AXIOM models object interactions, collisions, and coordinated behaviors between agents.
• Structured growth and model reduction
  Explains how AXIOM grows model complexity when needed and prunes it through Bayesian model reduction.
• Variational inference and data efficiency
  Describes how AXIOM uses variational free energy minimization to infer hidden causes efficiently without exhaustive computation.
• Planning for reward and information gain
  Explains how AXIOM balances goal achievement with actions that reduce uncertainty and improve future decisions.
• Performance and efficiency benchmarks
  Highlights AXIOM’s ability to outperform deep learning systems with dramatically less data and computation.
• Gaussian distributions and uncertainty modeling
  Introduces Gaussian distributions as a mathematical foundation for representing uncertainty in real-world environments.
• 3D Gaussian splatting
  Explains how Gaussian splatting enables real-time, continuously updated 3D scene representations.
• Variational Bayes Gaussian Splatting (VBGS)
  Describes how VBGS extends Gaussian splatting with variational inference to prevent forgetting and enable continuous learning.
• Uncertainty-aware navigation and exploration
  Shows how VBGS allows agents to balance exploration and exploitation by explicitly modeling uncertainty.
• VBGS beyond visual data
  Explains how VBGS can model uncertainty in non-visual domains such as finance, cybersecurity, maintenance, and sensor streams.
• Integration of AXIOM and VBGS
  Describes how AXIOM provides reasoning and decision-making while VBGS supplies real-time environmental modeling.
• Scaling to teams, enterprises, and cities
  Explores how these systems scale through distributed agents connected via the Spatial Web Protocol.
• The Spatial Web as connective tissue
  Explains how HSTP and HSML enable secure, interoperable belief sharing and coordination across global networks.
• Real-time digital twins
  Shows how AXIOM and VBGS enable living digital twins that update continuously rather than relying on static models.
• Enterprise and industrial applications
  Covers use cases including robotics, logistics, utilities, infrastructure monitoring, and smart environments.
• Explainability and ethical AI
  Explains why AXIOM’s Bayesian structure produces interpretable reasoning suitable for regulated and safety-critical domains.
• Preparing leaders for adoption
  Outlines why education, training, and guided activation are essential for organizations adopting Active Inference systems.