Motion-Based Autonomous Grounding: Inferring External World Properties from Encoded Internal Sensory States Alone

Yoonsuck Choe, Noah H. SmithAAAI 2006

Introduction: The Grounding Problem

In AI and robotics, the grounding problem is a fundamental hurdle: how does a system know that a specific internal state (like a pattern of neural firing or a bit-string) corresponds to a specific property in the real world (like a 45-degree edge)?

Usually, human programmers "cheat" by providing these semantics—we tell the robot that "Sensor A = 45-degree line." However, a truly autonomous agent "trapped" inside its own "brain" only sees internal spikes ($s$). It has no direct access to the external world ($I$) or the function ($f$) that turned the world into those spikes. For an agent to be truly autonomous, it must discover these meanings for itself.

Contrasting external and internal perspectives on sensory grounding.

Methods: Learning through Sensory Invariance

The authors propose a sensorimotor framework where the agent learns through interaction rather than passive observation.

  1. The Agent Model: The agent has a limited field of view and a set of orientation-tuned sensory primitives (modeled after primary visual cortex neurons). It can move its "gaze" in various directions.
  2. Sensory Invariance: The core hypothesis is that if a specific sensory unit is active (e.g., one that responds to a 45° orientation), moving the gaze along that 45° axis will keep that sensory state stable (invariant).
  3. Reinforcement Learning: The agent uses a standard Q-learning algorithm with a unique reward function. Instead of an external "Good job!" signal, the reward is simply the degree of invariance between the current sensory state and the previous one.
  4. Action Mapping: The agent learns a mapping from sensory states to motor actions that maximize this reward. If it succeeds, the resulting motor behavior "mirrors" the external stimulus property.
Maintaining sensory invariance through gaze movement along a stimulus axis.

Key Insights: Action as Semantics

The "meaning" of a sensory state isn't a label; it's the action required to maintain that state. If a sensation is grounded in a 45° diagonal movement, the agent "understands" the sensation as a 45° orientation. This feedback loop provides the necessary bridge to external reality, allowing an internal observer to "feel" the consequences of its own actions on its sensations. This approach aligns with biological development, where simple invariance-seeking behavior might organize early neural pathways.

Results and Success Metrics

Experiments were conducted using natural images (e.g., radial leaf patterns) pre-processed with a Difference-of-Gaussian (DoG) filter to simulate biological vision. The sensory system utilized arrays of 8, 16, or 32 oriented Gabor filters.

  • Q-Table Convergence: The learned Q-tables closely matched theoretical ideals, exhibiting a clear diagonal structure that paired specific orientations with their corresponding directions of motion.
  • Quantitative Success: The Root Mean Squared Error (RMSE) between the learned and ideal Q-tables converged to a low stable level across all tested sensor counts.
  • Qualitative Trajectories: Before learning, the agent’s gaze moved randomly. After learning, the gaze trajectories followed the prominent edges of the natural image (e.g., tracing the lines of the leaves), demonstrating that the agent had successfully "grounded" its internal states in the physical structure of the environment.

Limitations and Future Directions

While the current model successfully grounds simple properties like orientation, scaling to complex object concepts (like "a chair") remains a challenge. Future work may require long-term memory to integrate motor trajectories over time. Additionally, the authors suggest a link to "mirror neurons" in the brain, hypothesizing they might play a role in this kind of sensorimotor grounding. Future research will explore how these low-level motor primitives can serve as a foundation for higher-level object recognition.

Summary updated

BibTeX
@inproceedings{choe2006motion,
  author = {Yoonsuck Choe and Noah H. Smith},
  title = {{Motion-Based Autonomous Grounding: Inferring External World Properties from Encoded Internal Sensory States Alone}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2006},
}

Loading related content...

Figure