🤖 machine learning

Evidence of an Emergent "Self" in Continual Robot Learning

This paper proposes that the "self" in intelligent systems can be identified as an invariant cognitive subnetwork, demonstrating through continual robot learning experiments that robots exposed to variable tasks develop significantly more stable internal structures compared to those learning constant tasks.

Original authors: Adidev Jhunjhunwala, Judah Goldfeder, Hod Lipson

Published 2026-03-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Adidev Jhunjhunwala, Judah Goldfeder, Hod Lipson

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Question: Does a Robot Have a "Self"?

Imagine you drop two people into a swimming pool.

Person A has spent 20 years running, jumping, and dancing. They know exactly how their body moves, where their limits are, and how their muscles work together. Even though they don't know how to swim yet, they have a deep, internal sense of who they are physically.
Person B has never moved a muscle in their life. They have no idea how their body works.

When both try to learn to swim, Person A has a huge advantage. They aren't just learning "how to swim"; they are learning how to apply their existing knowledge of their own body to a new task. They have a "self."

The big question this paper asks: Can a robot develop this same kind of "self"? Can a robot learn a bunch of different tasks and, in the process, build a permanent, unchanging internal map of its own body, separate from the specific tricks it's learning?

The Experiment: The Robot Gym

The researchers set up a digital gym for a four-legged robot (a quadruped). They didn't just teach it one thing; they put it through a "continual learning" boot camp.

The Training: The robot had to learn three very different moves in a loop:
- Walk: Moving forward.
- Wiggle: Spinning in place.
- Bob: Jumping up and down.
The Switch: Every time the robot got good at one, they switched to the next. It had to forget the old trick and learn the new one, over and over again.
The Control Group: They also had a "control" robot that only learned to walk, over and over, for the same amount of time.

The Discovery: The "Core" vs. The "Costume"

When the researchers looked inside the robot's brain (its neural network), they found something amazing.

Think of the robot's brain like a theater production:

The "Costume" (Task-Specific Parts): These are the parts of the brain that change every time the robot learns a new trick. When the robot learns to "wiggle," these parts put on a "wiggle costume." When it learns to "bob," they swap it for a "bob costume." These parts are flexible and change constantly.
The "Actor" (The Self): This is the part of the brain that stays the same. It's the core understanding of the robot's body: I have four legs, I have joints here, I have this much weight.

What they found:

The "Walk-Only" Robot: Because it only ever did one thing, its brain was a messy jumble. It didn't need to separate "who I am" from "what I'm doing" because it never had to switch. Its brain was like a single, tangled knot.
The "Continual Learning" Robot: Because it had to switch between walking, wiggling, and bobbing, its brain naturally organized itself. It developed a stable core (the "Self") that stayed perfectly consistent, while the rest of the brain rearranged itself to handle the new tasks.

The Analogy: The Swiss Army Knife vs. The Toolbox

The Single-Task Robot is like a Swiss Army Knife where every tool is welded to the handle. If you need a screwdriver, you have to use the whole knife. It works, but it's rigid.
The Continual Learning Robot is like a high-end Toolbox. It has a sturdy, permanent handle (the "Self") that you hold onto no matter what. But the tools you clip onto it change. You clip on a hammer for walking, a saw for wiggling, and a drill for bobbing. The handle never changes; it knows exactly how to hold the tools.

Why This Matters

This is a huge deal for Artificial Intelligence. Usually, when we train AI, we treat it as a "black box" that just outputs answers. We don't know how it thinks.

This paper suggests that if you force an AI to learn many different things in a row, it naturally builds a "self." It creates a permanent, reusable understanding of its own existence (its body and physics) that it can carry over to new tasks.

The Takeaway:
You don't need to program a robot with a specific "self-awareness" module. If you just let it learn and adapt to a changing world, a "self" will emerge on its own. It's the part of the brain that says, "I am this body," while the rest of the brain figures out, "Okay, today I need to walk."

This gives us a new way to look at AI: not just as a collection of rules, but as a system that can develop a persistent identity through experience.

1. Problem Statement

The central challenge addressed is the lack of a principled method to quantify whether an intelligent system possesses a concept of a "self" and how to distinguish this "self" from other cognitive structures (e.g., task-specific skills).

The Hypothesis: The authors propose that the "self" can be isolated as the invariant portion of a cognitive process. While skills and knowledge change rapidly with new experiences, the internal model of one's own body (kinematics, dynamics, morphology) should remain persistent.
The Gap: Current deep reinforcement learning (RL) policies for robot locomotion are often treated as monolithic functions where internal representations are entangled. While some research explicitly separates self-modeling from controllers, it is unknown if a stable, reusable "self-model" can emerge spontaneously within a standard neural network trained across multiple behaviors without explicit architectural constraints.

2. Methodology

Experimental Setup

Robot: A simulated quadruped (Ant-like morphology) with 8 actuated degrees of freedom (2 joints per leg).
Learning Algorithm: Soft Actor-Critic (SAC) implemented in rl_games.
Network Architecture: 2-layer Multi-Layer Perceptron (MLP) with 150 hidden units per layer and ReLU activations.
Training Regimes:
1. Continual Learning (Experimental): The robot is trained sequentially on three distinct behaviors in a cyclical manner: Walk (forward translation), Wiggle (in-place rotation), and Bob (vertical hopping). The policy weights are transferred from one phase to the next.
2. Constant Task (Control): The robot is trained on a single behavior (e.g., Walk only) for an equivalent number of total training cycles to serve as a baseline.

Analysis Framework

To detect an emergent "self," the authors analyzed the internal structure of the trained neural networks using co-activation analysis:

Shared Reference States: All policies were evaluated on a shared set of input states spanning all behaviors to ensure fair comparison.
Co-Activation Matrix: For each hidden layer, a neuron-neuron cosine similarity matrix was computed based on activation traces across the shared states.
Block Diagonalization: The similarity matrix was thresholded and treated as a graph. Connected components (subnetworks) were identified using block diagonalization (Reverse Cuthill-McKee ordering).
Persistence Scoring: Neurons were aligned across different behavior policies (solving an optimal assignment problem). A persistence score was calculated for each neuron family based on:
- Activation Similarity: How similarly the neuron activates across behaviors.
- Connectivity Similarity: How similarly the neuron's co-activation pattern with other neurons is preserved.
Differentiation: The authors compared the stability of the largest subnetwork (hypothesized "self") against the pooled remainder of the network (hypothesized "task-specific" components).

3. Key Contributions

Emergent Self-Model: Demonstrated that a standard deep RL policy, when subjected to continual multi-behavior learning, spontaneously organizes into a stable, invariant subnetwork that persists across task switches, distinct from the plastic, task-specific components.
Quantitative Metric for "Self": Proposed a method to quantify "selfhood" in AI by measuring the persistence score of neural subnetworks. A high persistence score indicates a representation that is robust to distribution shifts (new tasks).
Differentiation from Monolithic Learning: Showed that this structure does not emerge in single-task training (even with extended duration), proving that the "self" is a specific byproduct of the tension between learning new skills and retaining body knowledge in continual learning.
Sim-to-Real Validation: Validated that the learned behaviors (Walk, Wiggle, Bob) are kinematically valid on a physical quadruped robot, ensuring the findings are not merely simulation artifacts.

4. Key Results

Stable Core vs. Plastic Periphery: In the continual learning condition, the largest subnetwork (the "self") exhibited significantly higher stability ( $p \ll 0.001$ $p ≪ 0.001$ ) compared to the rest of the network.
- Persistence Gap: The mean separation between the "self" subnetwork and the "task" subnetwork was 16.9 percentage points (99% confidence lower bound: 16.68).
- Visual Evidence: Visualization (Figure 2) showed that in continual learning, the "self" subnetwork forms a continuous, dominant band across Walk, Wiggle, and Bob policies. In contrast, the single-task baseline showed diffuse, fragmented connections with no dominant stable core.
Layer-Specific Emergence:
- Layer 1: The "self" subnetwork was most prominent, saturating at ~80 neurons (out of 150).
- Layer 2: A smaller but still identifiable stable core (~30 neurons) emerged, while the rest of the layer remained more behavior-dependent.
Robustness: The phenomenon was consistent across 10 independent random seeds and various architectural hyperparameters (width, depth, activation functions).
Control Baseline: In the "Walk-only" control, the network did not develop a distinct, large invariant core. The persistence scores were distributed smoothly, indicating that the "self" is not a result of over-training but specifically of multi-behavior switching.

5. Significance and Implications

Redefining Self-Awareness in AI: The paper suggests that "self" in AI need not be an explicit, separate module or a hand-coded model. Instead, it can be an emergent property of a neural network optimizing for efficiency across changing tasks. The "self" is the reusable, body-centric representation that allows the agent to adapt quickly to new tasks without catastrophic forgetting of its physical constraints.
Practical Applications:
- Modular Learning: Identifying the "self" subnetwork could allow robots to reuse this core as a backbone when learning entirely new skills, accelerating training.
- Damage Resilience: Monitoring the stability of this subnetwork could help detect when a robot's body schema is being corrupted (e.g., by physical damage) versus when it is simply learning a new skill.
- Continual Learning Strategies: The findings offer a diagnostic tool to distinguish between useful representation reuse and catastrophic interference in continual learning systems.
Future Directions: The authors suggest scaling this approach to more complex morphologies (humanoids) and broader task families to test if the "self" subnetwork remains stable and useful as the complexity of the agent's experience grows.

In summary, the paper provides empirical evidence that continual learning induces a modular architecture in deep RL policies, where a persistent "self" subnetwork emerges to encode the invariant dynamics of the robot's body, separating it from the transient, task-specific control logic.